<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3705">
  <Title>Language Engineering and the Pathway to Healthcare: A user-oriented view</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
3 Feedback and verification
</SectionTitle>
    <Paragraph position="0"> Translation accuracy is of course crucial in the medical domain, and sometimes problematic even with human interpreters, if not trained properly (Flores, 2005). Both speech recognition (SR) and translation are potential sources of error in the SLT chain, so it is normal and necessary to incorporate in SLT systems the provision of feedback and verification for users. The standard method for SR is textual representation, often in the form of a list of choices, for example as in Figure 3, from Precoda et al. (2004).</Paragraph>
    <Paragraph position="1">  particular design of the machine translation (MT) component (e.g. use of an interlingua representation, as in MedSLT, Speechalator). In the Transonics system, the SR accuracy is automatically assessed by the MT component: SR output that conforms to the expectations of the MT systems grammar is preferred. null For the literate English-speaking user, this approach seems reasonable, although an interface such as the one shown in Figure 4, detailing the output of the parse must be of limited utility to a doctor with no linguistics training, and we must assume that the prototype is designed more for the developers' benefit than for the end-users.</Paragraph>
    <Paragraph position="2"> For the patient with limited or no English, the issue of feedback and verification is much more difficult. As mentioned above, and reiterated by Precoda et al. (2004), the user may not be (wholly) literate, or indeed the language (or dialect) may not have an established writing system. For some languages, displaying text in the native orthography may be an added burden. Figure 5 shows Speechalator's Arabic input screen (Waibel et al., 2003). It is acknowledged that the users must &amp;quot;know something about the operation of the machine&amp;quot;, and although it is stated that the display uses the writing system of the language to be recognised, in the illustration the Arabic is shown in transcription.</Paragraph>
    <Paragraph position="3"> Another issue concerns the ease with which a lay user can make any sense of a task in which they are asked to judge a number of paraphrases, some ungrammatical. This is an intellectual task that is difficult for someone with limited education or no experience of linguistic &amp;quot;games&amp;quot;. For example, for this reason we have rejected the use of semantically unpredictable sentences (SUS) (Beno^it et al., 1996) in our attempts to evaluate Somali speech synthesis (Somers et al., 2006). This leads us to a consideration of how medical SLT can best be evaluated.</Paragraph>
  </Section>
class="xml-element"></Paper>