<?xml version="1.0" standalone="yes"?>
<Paper uid="H92-1060">
  <Title>A Relaxation Method for Understanding Speech Utterances 1</Title>
  <Section position="6" start_page="303" end_page="303" type="concl">
    <SectionTitle>
CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> Through examining a large body of speech material collected from a general population of naive users, we have reached the conclusion that it is not feasible to design a grammar that can always achieve a complete linguistic analysis of every input sentence. We have simultaneously become aware that a system that could recover a partial analysis would also be valuable for overcoming some recognition errors. We have described in this paper a capability to produce a partial analysis whenever a full parse fails, and have reported substantial performance improvements on test material as a direct consequence of this robust mechanism. We were able to leverage off of existing system components to a large extent, leading to a rapid development of the new robust parsing mechanism. This capability allowed the system to answer many more sentences than had previously been possible.</Paragraph>
    <Paragraph position="1"> We have begun to explore some possibilities for making use of a set of N-best recognizer outputs, by parsing a network of paths generated through an intelligent join of the top-N candidates. We can use the frequency of occurrence of a word in the top-N candidates as a measure of its robustness, and then select a path through the network that maximizes the selection of linguistically meaningful phrases that recurred among the top-N sentences. null We have just begun to incorporate robust parsing into our data-collection procedure. We have collected data for 4 scenarios from each of 15 subjects, where the system was toggled between robust and non-robust modes half way through each subject's episode. Subjects were asked to solw~ the scenarios, all of which had a unique answer.</Paragraph>
    <Paragraph position="2"> Interestingly, subjects were able to find the correct answer in robust mode 90% of the time, v.s. only 70% in the non-robust mode. We take this as a clear indicator that robust mode is effective in real usage. For a further discussion of this experiment see \[11\].</Paragraph>
  </Section>
class="xml-element"></Paper>