<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1040">
  <Title>Robust parsing of severely corrupted spoken utterances</Title>
  <Section position="6" start_page="197" end_page="199" type="evalu">
    <SectionTitle>
5 Experimental results
</SectionTitle>
    <Paragraph position="0"> The above ideas have been implemented in a parser called SYNAPSIS (from SYNtax-Aided Parser for Semantic Interpretation of Speech). SYNAPSIS is an evolution of the parser included in the SUSY system for understanding speech and described in \[Poesio 87\]. SYNAPSIS has been implemented in Common Lisp and relies on about 150 KSs, able to handle a 1011-word lexicon on a restricted semantic domain. An idea of the lingulstic coverage is given by the equivalent branching factor, which is  tices generated by processing as many sentences uttered in continuous speech with natural intonation in a normal office environment. The overall performance results in about 80% correct sentence understanding \[Fissore 88\].</Paragraph>
    <Paragraph position="1"> The thresholds for JVERIFY have been experimentally determined to minimize the computational load, represented by the average number of Dis generated during each parsing. Tab. 1 shows the number of jolly words that have been skipped by the parser vs. the number of jollies actually missing in the corresponding lattices. The former figures are higher than the latter, indicating that many words, albeit present, have been discarded by JVERIFY because of their bad acoustical scores or their scarce contribution to contraint propagation.</Paragraph>
    <Paragraph position="2"> The most apparent advantage of the above technique is the increase in the number of sentences that can be analyzed without querying the user for lacking information. Tab. 2 displays the number of lattices, corresponding to the sentences containing at least one word of jolly type, in which some of such words are missing. It is seen that about 75~ of them have been successfully understood.</Paragraph>
    <Paragraph position="3"> This figure does not change substantially as the number of missing jollies per sentence increases, and hence indicates robustness. The computational load, given by the number of generated Dis, is somewhat affected by the number of missing jollies. However, this is mainly due to the fact that sentences with many jollies are also longer  and syntactically complex. The actual efficiency can be better estimated from Fig. 4, where the average number of generated Dis is plot as a function of the threshold on the width of the jolly temporal 'hole'. The figure displays also the amount of parsing failures related to jolly problems (failures due to other reasons have been ignored for simplicity). The curve indicates that raising the threshold does not change much the number generated Dis (the relative oscillations of the values are small). This means that the relaxation of constraints during the application of JVERIFY is not a source of inefficiency. Moreover, there is a large range of values for which the parsing failure remains low.</Paragraph>
    <Paragraph position="4"> The curve also shows that relaxing constraints may even speed up the parsing. This can be easily explained. When the threshold is low, no jolly is skipped, and failure occurs when jollies are missing from the lattice. When the threshold is raised, skipping begins to work: good-scored false jollies are no more a source of disturbance, and correct but bad-scored jollies are skipped thus avoiding to delay the parsing; as a consequence the overall number of Dis decreases. Further enlarging the threshold reverts this tendency, since the too-much-relaxed constraints allow the aggregation of words that would have been discarded with stricter constraints; failures occur when one of such aggregations makes up a complete parse scoring</Paragraph>
  </Section>
class="xml-element"></Paper>