<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-3164">
  <Title>A SPOKEN LANGUAGE TRANSLATION SYSTEM : SL-TRANS2</Title>
  <Section position="8" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
6. Experiment
</SectionTitle>
    <Paragraph position="0"> The SL-TKANS2 system as developed so far can treat utterances from about 600 vocabulary items. It runs on a UNIX-workstation such as SPARC-2. Average run time for one utterance is about one minute, half for speech recognition and half for translation. A preliminary experiment has been carried out on five dialogues, which include 101 sentences. The results are summarized in Table 2. Input speech material are specific speaker's utterances. Abilities oPS  speech recognition (SR) and translation (TR) as a single component arc about 85% and 98% respectively. Correctness of translation is evaluated whether the generated sentences arc grammatically correct and semantically understandable (minor errors those involving determiners are ignored). We can see that the translation part has high translation ability as a *3) The xperiment was carried out on string inputs ACRES DE COLING-92, NANTES, 23-28 AOI~T 1992 1 0 5 1 PROC, OF COLING-92, NANTES, AUG. 23-28, 1992 single component. Only two sentences indicated below fall to translate properly.</Paragraph>
    <Paragraph position="1">  E2: Please see itsince I send an announcement of the conference In J1, the subject of the sentence is not uttered explicitly, because it is easily inferable as &amp;quot;registration-fee&amp;quot; from the context. The system, however, erroneously supplements 'T' since no honorific expression appears in the sentence.</Paragraph>
    <Paragraph position="2"> In Japanese sentence J2, a main clause appears later and a pronoun &amp;quot;sore&amp;quot; (it) referring to &amp;quot;teuroku-youshi&amp;quot; (a registration-form} in the first clause is used. The system does not see this referential relationship, and so it fails to generate a proper English sentence.</Paragraph>
    <Paragraph position="3"> As a total system, about 84% of the sample sentences are recognized and translated into English correctly. Some examples are shown below.</Paragraph>
    <Paragraph position="4"> J3: 9-gatsu 27-niehi-ikou-no torikeshi-nitaisuru haraimodoshi-wa dekimasen E3: The refund to the cancellation after September 27th is not possible.</Paragraph>
    <Paragraph position="5"> J4: dewa dareka-ga watashi-no kawari-ni sankasuru-koto-wa dekimasu-ka? E4: Then, can someone attend instead of me? Generally speaking, it is desirable that a translation system be able to detect erroneous speech recognition output. In our system, most of such failures are filtered out, but two sentences are translated into English. These undesirable outcomes are due to inadequacy of selectional restrictions used in the translation component, as indicated below.</Paragraph>
  </Section>
class="xml-element"></Paper>