<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1013">
  <Title>Progress Report on the Chronus System: ATIS Benchmark Results</Title>
  <Section position="10" start_page="70" end_page="70" type="concl">
    <SectionTitle>
6. CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> In this paper we give a global outline of the CHRONUS speech understanding system. The system is built around the conceptual decoder, a Viterbi decoder that uses a stochastic model for extracting the conceptual content of an input sentence. Although the problem is formalized in such a way that the decoder could also extract the actual value of the conceptual relations (not only their category), the limited amount of training sentences currently available suggested the use of a more traditional pattern matcher (the template generator) along with the conceptual decoder. However, the advantage of the stochastic formalization is the trainability of the model over a database of suitably annotated examples. The annotation given with the MADCOW sentences and the comparator developed by NIST provide a useful feedback signal that allows to automatize the training procedure. In a preliminary experiment designed to test this procedure we show that a significant improvement of the accuracy of the system can be obtained without human supervision.</Paragraph>
    <Paragraph position="1"> The results on the February 92 ATIS test are then reported in the paper. The big discrepancy between the official NIST score and the score obtained in a successive assessment of the system is explained by inaccura-</Paragraph>
  </Section>
class="xml-element"></Paper>