<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1013">
  <Title>Progress Report on the Chronus System: ATIS Benchmark Results</Title>
  <Section position="3" start_page="67" end_page="67" type="intro">
    <SectionTitle>
2. SYSTEM ARCHITECTURE
</SectionTitle>
    <Paragraph position="0"> The task of the conceptual decoder (see Fig. I) is that of providing a conceptual segmentation</Paragraph>
    <Paragraph position="2"> of words that express the concept relation cj within the given sentence In the current version of the CHRONUS understanding system the speech recognizer is used in a decoupled mode. The best string of words produced by the recognizer is used by the decoder for generating the conceptual segmentation. Because in this particular task there are numbers, acronyms and compound words, the string is pre-processed by a module called lezical parser that generates a lattice with all the possible interpretations of the string (e.g. the substring &amp;quot;B SEVEN FOUR SEVEN&amp;quot; could be interpreted as &amp;quot;B 747&amp;quot; or &amp;quot;B7 47&amp;quot; or &amp;quot;B74 7&amp;quot;, etc. The conceptual decoder is then realized as a generalization of the Viterbi algorithm that works on a lattice rather than on a string of words.</Paragraph>
    <Paragraph position="3"> The template generator \[6\] consists of a simple pattern matching procedure that, given the conceptual segmentation, produces for each concept relation cj the corresponding concept value vj. Finally the SQL translator translates the meaning representation M into an SQL query.</Paragraph>
  </Section>
class="xml-element"></Paper>