<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-2034">
  <Title>A speech interface for open-domain question-answering</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Related research
</SectionTitle>
    <Paragraph position="0"> Kupiec and others (1994) at Xerox labs built one of the earliest spoken information retrieval systems, with a speaker-dependent isolated-word speech recognizer and an electronic encyclopedia. One reason they reported for the success of their system was their use of simple language models to exploit the observation that pairs of words co-occurring in a document source are likely to be spoken together as keywords in a query. Later research at CMU built upon similar intuition by deriving the language-model of their Sphinx-II speech recognizer from the searched document source. Colineau and others (1999) developed a system as a part of the THISL project for retrieval from broadcast news to respond to news-related queries such as What do you have on . . . ? and I am doing a report on . . . -- can you help me? The queries the authors addressed had a simple structure, and they successfully modelled them in two parts: a question-frame, for which they handwrote grammar rules; and a content-bearing string of keywords, for which they fitted standard lexical language-models from the news collection.</Paragraph>
    <Paragraph position="1"> Extensive research (Garofolo et al., 2000; Allan, 2001) has concluded that spoken documents can be effectively indexed and searched with word-error rates as high as 30-40%. One might expect a much higher sensitivity to recognition errors with a short query or natural-language question. Two studies (et al., 1997; Crestani, 2002) have measured the detrimental effect of speech recognition errors on the precision of document retrieval and found that this task can be somewhat robust to 25% word-error rates for queries of 2-8 words.</Paragraph>
    <Paragraph position="2"> Two recent systems are worthy of special mention. First, Google Labs deployed a speaker-independent system in late 2001 as a demo of a telephone-interface to its popular search engine. (It is still live as of April 2003.) Second, Chang and others (2002a; 2002b) have implemented systems for the Pocket PC that interpret queries spoken in English or Chinese. This last group appears to be at the forefront of current research in spoken interfaces for document retrieval.</Paragraph>
    <Paragraph position="3"> None of the above are question-answering systems; they boil utterances down to strings of keywords, discarding any other information, and return only lists of matching documents. To our knowledge automatic answering of spoken natural-language questions has not previously been attempted.</Paragraph>
  </Section>
class="xml-element"></Paper>