<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1056">
  <Title>Using Conditional Random Fields For Sentence Boundary Detection In Speech</Title>
  <Section position="9" start_page="456" end_page="457" type="concl">
    <SectionTitle>
5 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> Finding sentence boundaries in speech transcriptions is important for improving readability and aiding downstream language processing modules. In this paper, prosodic and textual knowledge sources are integrated for detecting sentence boundaries in speech. We have shown that a discriminatively trained CRF model is a competitive approach for the sentence boundary detection task. The CRF combines the advantages of being discriminatively trained and able to model the entire sequence, and so it outperforms the HMM and Maxent approaches  consistently across various testing conditions. The CRF takes longer to train than the HMM and Max-ent models, especially when the number of features becomes large; the HMM requires the least training time of all approaches. We also find that as more features are used, the differences among the modeling approaches decrease. We have explored different approaches to modeling various knowledge sources in an attempt to achieve good performance for sentence boundary detection. Note that we have not fully optimized each modeling approach. For example, for the HMM, using discriminative training methods is likely to improve system performance, but possibly at a cost of reducing the accuracy of the combined system.</Paragraph>
    <Paragraph position="1"> In future work, we will examine the effect of Viterbi decoding versus forward-backward decoding for the CRF approach, since the latter better matches the classification accuracy metric. To improve SU detection results on the STT condition, we plan to investigate approaches that model recognition uncertainty in order to mitigate the effect of word errors. Another future direction is to investigate how to effectively incorporate prosodic features more directly in the Maxent or CRF framework, rather than using a separate prosody model and then binning the resulting posterior probabilities.</Paragraph>
    <Paragraph position="2"> Important ongoing work includes investigating the impact of SU detection on downstream language processing modules, such as parsing. For these applications, generating probabilistic SU decisions is crucial since that information can be more effectively used by subsequent modules.</Paragraph>
  </Section>
class="xml-element"></Paper>