<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3009">
  <Title>Using Higher-level Linguistic Knowledge for Speech Recognition Error Correction in a Spoken Q/A Dialog</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion and Future Works
</SectionTitle>
    <Paragraph position="0"> We proposed an improved syllable-based noisy channel model and combined higher level linguistic knowledge for semantic-oriented approach in a speech recognition error correction, which shows a superior performance in domain-specific IR applications.</Paragraph>
    <Paragraph position="1"> The previous works only focused on inter-word level error correction, commonly depending on a large amount of training corpus for the error correction model and the language model. So, previous approaches require enormous results of ASR and are dependent on specific speakers and environments. On the other hand, our method takes in far smaller training corpus, and it is possible to implement the method easily and in a short time to obtain the better error correction rate because it utilizes the semantic information of the application domain.</Paragraph>
    <Paragraph position="2"> And our semantic-oriented approach has more advantages over lexical based ones, since it is less sensitive to each error pattern. Also, the approach has a broader coverage of error patterns, since several similar common error strings in the semantic ground can be reduced to one semantic error pattern, which enables us to improve the probability of recovering from erroneous recognition results. null And, because the LSP scheme transforms pure lexical entries into abstract semantic categories, the size of the error pattern database can be reduced remarkably, and it also increases the coverage and robustness compared with the previous pure lexical entries that can only deal with the morphological variants.</Paragraph>
    <Paragraph position="3"> With all these facts, the LSP correction has a high possibility of generating semantically correct correction due to the massive use of semantic contexts. Hence, it shows a high performance, especially when combined with domain-specific speech-driven natural language IR and QA systems.</Paragraph>
    <Paragraph position="4"> Future work should include the end-performance experiments with IR or QA application for our error correction model.</Paragraph>
  </Section>
class="xml-element"></Paper>