<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1055">
  <Title>Position Specific Posterior Lattices for Indexing Speech</Title>
  <Section position="8" start_page="449" end_page="449" type="concl">
    <SectionTitle>
7 Conclusions and Future work
</SectionTitle>
    <Paragraph position="0"> We have developed a new representation for ASR lattices -- the Position Specific Posterior Lattice (PSPL) -- that lends itself naturally to indexing speech content and integrating state-of-the-art IR techniques that make use of proximity and context information. In addition, the PSPL representation is also much more compact at no loss in WER -- both 1-best and ORACLE.</Paragraph>
    <Paragraph position="1"> The retrieval results obtained by indexing the PSPL and performing adequate relevance ranking are 20% better than when using the ASR 1-best output, although still far from the performance achieved on text data.</Paragraph>
    <Paragraph position="2"> The experiments presented in this paper are truly a first step. We plan to gather a much larger number of queries. The binary relevance judgments -- a given document is deemed either relevant or irrelevant to a given query in the reference &amp;quot;ranking&amp;quot; -assumed by the standard trec_eval tool are also a serious shortcoming; a distance measure between rankings of documents needs to be used. Finally, using a baseline engine that in fact makes use of proximity and context information is a priority if such information is to be used in our algorithms.</Paragraph>
  </Section>
class="xml-element"></Paper>