<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-2136">
  <Title>Confirmation in Multimodal Systems</Title>
  <Section position="8" start_page="827" end_page="827" type="evalu">
    <SectionTitle>
6 DISCUSSION
</SectionTitle>
    <Paragraph position="0"> There are two likely reasons why late confLrmation outperforms early confLrmation: implicit confirmation and multirnodal disambiguation. Heisterkamp theorized that implicit confLrmation could reduce the number of turns in dialogue \[25\]. Rudnicky proved in a speech-only digit-entry system that implicit confirmation improved throughput when compared to explicit confirmation \[27\], and our results confirm their findings. Lavie and colleagues have shown the usefulness of late-stage disambiguafion, during which speechunderstanding systems pass multiple interpretations through the system, using context in the final stages of processing to disambiguate the recognition hypotheses \[28\]. However, we have demonstrated and empirically shown the advantage in combining these two strategies in a multirnodal system.</Paragraph>
    <Paragraph position="1"> It can be argued that implicit confirmation is equivalent to being able to undo the last command, as some multimodal systems allow \[3\]. However, commands that are infeasible, profound, risky, costly, or irreversible are difficult to undo. For this reason, we argue that implicit confirmation is often superior to the option of undoing the previous command. Implicit confirmation, when combined with late confirmation, contributes to a smoother, faster, and more accurate collaboration between human and computer.</Paragraph>
  </Section>
class="xml-element"></Paper>