<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3702">
  <Title>Marianne.Starlander@eti.unige.ch</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Medical applications have emerged as one of the most promising application areas for spoken language translation, but there is still little agreement about the question of architectures. There are in particular two architectural dimensions which we will address: general processing strategy (statistical or grammar-based), and top-level translation functionality (unidirectional or bidirectional translation). Given the current state of the art in recognition and machine translation technology, what is the most appropriate combination of choices along these two dimensions? Reflecting current trends, a common approach for speech translation systems is the statistical one. Statistical translation systems rely on parallel corpora of source and target language texts, from which a translation model is trained. However, this is not necessarily the best alternative in safety-critical medical applications. Anecdotally, many doctors express reluctance to trust a translation device whose output is not readily predictable, and most of the speech translation systems which have reached the stage of field testing rely on various types of grammar-based recognition and rule-based translation (Phraselator, 2006; S-MINDS, 2006; MedBridge, 2006). Even though statistical systems exhibit many desirable properties (purely datadriven, domain independence), grammar-based systems utilizing probabilistic context-free grammar tuning appear to deliver better results when training data is sparse (Rayner et al., 2005a).</Paragraph>
    <Paragraph position="1"> One drawback of grammar-based systems is that out-of-coverage utterances will be neither recognized nor translated, an objection that critics have sometimes painted as decisive. It is by no means obvious, however, that restricted coverage is such a serious problem. In text processing, work on several generations of controlled language systems has developed a range of techniques for keeping users within the bounds of system coverage (Kittredge, 2003; Mitamura, 1999). If these techniques work for text processing, it is surely not inconceivable that variants of them will be equally successful for spoken language applications. Users are usually able to adapt to a controlled language system given enough time. The critical questions are how to provide efficient support to guide them towards the system's coverage, and how much time they will then need before they have acclimatized.</Paragraph>
    <Paragraph position="2"> With regard to top-level translation functionality, the choice is between unidirectional and bidirectional systems. Bidirectional systems are certainly possible today1, but the arguments in favor of them are not as clear-cut as might first appear. Ceteris paribus, doctors would certainly prefer bidirectional systems; in particular, medical students are trained to conduct examination dialogues using &amp;quot;open questions&amp;quot; (WH-questions), and to avoid leading the patient by asking YNquestions. null The problem with a bidirectional system is, however, that open questions only really work well if the system can reliably handle a broad spectrum of replies from the patients, which is overoptimistic given the current state of the art. In practice, the system's coverage is always more or less restricted, and some experimentation is required before the user can understand what language it is capable of handling. A doctor, who uses the system regularly, will acquire the necessary familiarity.</Paragraph>
    <Paragraph position="3"> The same might be true for a few patients, if special circumstances mean that they encounter speech translation applications reasonably frequently. Most patients, however, will have had no previous exposure to the system, and may be unwilling to use a type of technology which they have trouble understanding.</Paragraph>
    <Paragraph position="4"> A unidirectional system, in which the doctor mostly asks YN-questions, will never be ideal. If, 1 For example, the S-MINDS system (S-MINDS, 2006) offers bidirectional translation.</Paragraph>
    <Paragraph position="5"> however, the doctor can become proficient in using it, it may still be very much better than the alternative of no translation assistance at all.</Paragraph>
    <Paragraph position="6"> To summarize, today's technology definitely lets us build unidirectional grammar-based medical speech translation systems which work for regular users who have had time to adapt to their limitations. While bidirectional systems are possible, the case for them is less obvious, since users on the patient side may not in practice be able to use them effectively.</Paragraph>
    <Paragraph position="7"> In this paper, we will empirically investigate the ability of medical students to adapt to the coverage of unidirectional spoken language translation system. We report a series of experiments, carried out using a French to English speech translation system, in which medical students with no previous experience to the system were asked to use it to carry out a series of verbal examinations on subjects who were simulating the symptoms of various types of medical conditions. Evaluation will be focused on usability. We primarily want to know how quickly subjects learn to use the system, and how their performance compares to that of expert users.</Paragraph>
  </Section>
class="xml-element"></Paper>