<?xml version="1.0" standalone="yes"?>
<Paper uid="J99-3003">
  <Title>Lucent Technologies Bell Laboratories</Title>
  <Section position="7" start_page="386" end_page="386" type="concl">
    <SectionTitle>
7. Conclusions
</SectionTitle>
    <Paragraph position="0"> We described and evaluated a domain-independent, automatically trained call router that takes one of three actions in response to a caller's request. It can route the call to a destination within the call center, attempt to dynamically formulate a disambiguation query, or route the call to a human operator. The routing module selects a set of candidate destinations based on n-gram terms extracted from the caller's request and a vector-based comparison between these n-gram terms and each possible destination.</Paragraph>
    <Paragraph position="1"> If disambiguation is necessary, a yes-no question or a wh-question is dynamically generated from among n-gram terms automatically extracted from the training data based on closeness, relevance, and disambiguating power. This query formulation process allows the system to tailor the disambiguating query to the caller's original request and the candidate destinations.</Paragraph>
    <Paragraph position="2"> We have further demonstrated the effectiveness of our call router by evaluating the call router on both transcriptions of caller requests and the output of an automatic speech recognizer on these requests. When the input to the call router is free of recognition errors, our system correctly routes 93.8% of the calls after redirecting 10.2% of all calls to a human operator. When using the output of a speech recognizer with an approximately 23% word error rate, the rejection rate drops to 9.3%, the upper bound of the router performance drops from 97.2% to 92.5%, and the lower bound of the performance drops from 75.6% to 72.2%, illustrating the robustness of our call router in the face of speech recognition errors.</Paragraph>
  </Section>
class="xml-element"></Paper>