<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3007">
  <Title>Robustness Issues in a Data-Driven Spoken Language Understanding System</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Spoken language is highly variable as different people use different words and sentence structures to convey the same meaning. Also, many utterances are grammaticallyincorrect or ill-formed. It thus remains an open issue as to how to provide robustness for large populations of non-expert users in spoken dialogue systems. The key component of a spoken language understanding (SLU) system is the semantic parser, which translates the users' utterances into semantic representations. Traditionally, most semantic parser systems have been built using hand-crafted semantic grammar rules and so-called robust parsing (Ward and Issar, 1996; Seneff, 1992; Dowding et al., 1994) is used to handle the ill-formed user input in which word patterns corresponding to semantic tokens are used to fill slots in different semantic frames in parallel. The frame with the highest score then yields the selected semantic representation.</Paragraph>
    <Paragraph position="1"> Formally speaking, the robustness of language (recognition, parsing, etc.) is a measure of the ability of human speakers to communicate despite incomplete information, ambiguity, and the constant element of surprise (Briscoe, 1996). In this paper, two aspects of SLU system performance are investigated: noise robustness and adaptability to different applications. For the former, we expect that an SLU system should maintain acceptable performance when given noisy input speech data. This requires, the understanding components of the SLU system to be able to correctly interpret the meaning of an utterance even when faced with recognition errors. For the latter, the SLU system should be readily adaptable to a different application using a relatively small set (e.g.</Paragraph>
    <Paragraph position="2"> less than 100) of adaptation utterances.</Paragraph>
    <Paragraph position="3"> The rest of the paper is organized as follows. An overview of our data-driven SLU system is outlined in section 2. Experimental results on performance under a range of SNRs are then presented in section 3. Section 4 discusses adaptation of the HVS model to new applications. Finally, section 5 concludes the paper.</Paragraph>
  </Section>
class="xml-element"></Paper>