<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3002">
  <Title>Hybrid Statistical and Structural Semantic Modeling for Thai Multi- Stage Spoken Language Understanding</Title>
  <Section position="6" start_page="75" end_page="75" type="concl">
    <SectionTitle>
6 Conclusion and Future Works
</SectionTitle>
    <Paragraph position="0"> Recently, a multi-stage spoken language understanding (SLU) approach has been proposed for the first Thai spoken dialogue system. This article reported an improvement on the SLU system by replacing the regular grammar-based semantic model by a hybrid n-gram and regular grammar approach, which not only captures long-distant dependencies of word syntax, but also provides robustness against speech-recognition errors. The proposed model, called logical n-gram modeling, obviously improved the performance in every SLU stage, while reducing the computational time compared to the original regular-grammar approach. Under the probabilistic WFST framework, the system was improved further by using N-best wordhypotheses from the ASR, requiring only a small additional processing time compared to the use of 1-best. Further improvement of overall speech understanding as well as a spoken dialogue system in the future can be expected by introducing dialogue-state dependent modeling in the ASR and/or the SLU. A better way to utilize the first P-best goal hypotheses produced by the goal identifier instead of 1-best would also enhance the understanding performance.</Paragraph>
  </Section>
class="xml-element"></Paper>