<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2019">
  <Title>Early Deletion of Fillers In Processing Conversational Speech</Title>
  <Section position="2" start_page="0" end_page="73" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> This paper evaluates the benefit of deleting fillers early in parsing conversational speech. We follow LDC (2004) conventions in using the term filler to encompass a broad set of vocalized space-fillers that can introduce syntactic (and semantic) ambiguity.</Paragraph>
    <Paragraph position="1"> For example, in the questions Did you know I do that? Is it like that one? colloquial use of fillers, indicated below through use of commas, can yield alternative readings Did, you know, I do that? Is it, like, that one? Readings of the first example differ in querying listener knowledge versus speaker action, while readings of the second differ in querying similarity versus exact match. Though an engaged listener rarely has difficulty distinguishing between such alternatives, studies show that deleting disfluencies from transcripts improves readability with no reduction in reading comprehension (Jones et al., 2003).</Paragraph>
    <Paragraph position="2"> The fact that disfluencies can be completely removed without compromising meaning is important. Earlier work had already made this claim regarding speech repairs1 and argued that there was consequently little value in syntactically analyzing repairs or evaluating our ability to do so (Charniak and Johnson, 2001). Moreover, this work showed that collateral damage to parse accuracy caused by repairs could be averted by deleting them prior to parsing, and this finding has been confirmed in subsequent studies (Kahn et al., 2005; Harper et al., 2005). But whereas speech repairs have received significant attention in the parsing literature, fillers have been relatively neglected. While one study has shown that the presence of interjection and parenthetical constituents in conversational speech reduces parse accuracy (Engel et al., 2002), these constituent types are defined to cover both fluent and disfluent speech phenomena (Taylor, 1996), leaving the impact of fillers alone unclear.</Paragraph>
    <Paragraph position="3"> In our study, disfluency annotations (Taylor, 1995) are leveraged to identify fillers precisely, and these annotations are merged with treebank syntax. Extending the arguments of Charniak and Johnson with regard to repairs (2001), we argue there is little value in recovering the syntactic structure  of fillers, and we relax evaluation metrics accordingly (SS3.2). Experiments performed (SS3.3) use a state-of-the-art parser (Charniak, 2000) to study the impact of early filler deletion under in-domain and out-of-domain (i.e. adaptation) training conditions.</Paragraph>
    <Paragraph position="4"> In terms of adaptation, there is tremendous potential in applying textual tools and training data to processing transcribed speech (e.g. machine translation, information extraction, etc.), and bleaching speech data to more closely resemble text has been shown to improve accuracy with some text-based processing tasks (Rosenfeld et al., 1995). For our study, a state-of-the-art filler detector (Johnson et al., 2004) is employed to delete fillers prior to parsing.</Paragraph>
    <Paragraph position="5"> Results show parse accuracy improves significantly, suggesting disfluency filtering may have a broad role in enabling text-based processing of speech data.</Paragraph>
  </Section>
class="xml-element"></Paper>