<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1409">
  <Title>An Extended Architecture for Robust Generation*</Title>
  <Section position="3" start_page="63" end_page="64" type="intro">
    <SectionTitle>
2 Task-inherent and Technical
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="63" end_page="63" type="sub_section">
      <SectionTitle>
Problems
</SectionTitle>
      <Paragraph position="0"> The problems for generation that arise in a speech-to-speech translation system fall into two main classes: as in any large-scale system, there will be software-engineering problems which we will call technical problems and there are task-inherent problems that are particular to the translation task and the highly variable input in spontaneous speech.</Paragraph>
      <Paragraph position="1"> Since it is impossible to draw a clear line between technical and task-inherent problems, we will present a short classification and then go into more detail without much discussion whether a certain problem should be viewed as technical or taskinherent. null One would hope to be able to eliminate technical problems completely. However, in a large system, where development is distributed over many modules (implemented at different sites), some robustness against certain technical problems can become a necessity, as our experiences have shown. This is even more important during the development phasewhich a research system never leaves. Most technical problems have to do with violations of the interface definitions. Thisranges. from simple ~things such as using unknown predicates in the semantic representation to complex constructions that cannot be generated (the generation gap). We actually regard the latter as a task-inherent problem.</Paragraph>
      <Paragraph position="2"> Secondly, tile task-inherent problems can be divided into problems that are caused by (i) spontaneous speech input and (ii) insufficiencies in the analysis and translation steps.</Paragraph>
    </Section>
    <Section position="2" start_page="63" end_page="63" type="sub_section">
      <SectionTitle>
2.1 Robustness in Analysis
</SectionTitle>
      <Paragraph position="0"> The problems in (i) are quite varied and many cases are dealt with in analysis (and translation), some cases are dealt with in our robustness preprocessing submodule, a few in the classical submodules of generation. For example, there is a separate module on the level of speech recognition which deals with hesitations and self-corrections. Phenomena like ellipsis, phrasal and other incomplete utterances are handled by analysis, so generation must be able to deal with the corresponding semantic representations too. Agreement errors are handled (i.e., corrected) in analysis. But some serious syntactic errors cannot be corrected. However, at least the maximal analyzable segments are determined so that ungrammatical utterances are translated as sequences of several meaningful segments.</Paragraph>
    </Section>
    <Section position="3" start_page="63" end_page="64" type="sub_section">
      <SectionTitle>
2.2 Robustness in Generation
</SectionTitle>
      <Paragraph position="0"> The problems in (ii) are caused by an accunmlation of problems which result in (semantic) input to the generator that cannot be processed. Robustness in our system concentrates on this type of problenl which is and should be handled as a separate step between analysis/transfer and generation. (See the discussion of the architecture in section 3.) The list below contains some examples that are picked up again in section 4.</Paragraph>
      <Paragraph position="1">  * Problems with the structure of the semantic representation: null - unconnected subgraphs - multiple predicates referring to the same object  - omission of obligatory arguments (r) Problems with the content of the semantic representation: null - contradicting information - missing information (e.g. agreement information) null</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>