<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1055">
  <Title>Training Vocab. Unknown</Title>
  <Section position="4" start_page="285" end_page="286" type="concl">
    <SectionTitle>
CONCLUSION
</SectionTitle>
    <Paragraph position="0"> Stochastic language models will be an important component in future speech recognition and understanding systems. The N-gram language models are a class of model which is easily trained from observed data and provides  significant constraints to the recognition process and were therefore chosen for use in the informal ATIS CSR base-line evaluation test. The required number of parameters, however, is too large to be trained from practical amounts of data. Backing off to lower order models to estimate the probability of unobserved N-grams is an effective method for dealing with finite training data. The fact that these models are purely data driven is both an advantage and a disadvantage--they are free from often erroneous human bias, but also cannot incorporate human knowledge. One method of incorporating human knowledge in limited tasks is to smooth the probability estimates by grouping the words into human-defined classes and estimating the language model on the classes.</Paragraph>
    <Paragraph position="1"> The stack decoder is an attractive control strategy for a speech understanding system because it can combine information from the acoustic matching and any of a variety of language models/natural language systems into a single integrated search. The current prototype is not mature enough to use in a practical recognition/understanding system, but is showing promise. The no-grammar recognition works fairly well--but no-grammar recognition is not the goal. The goal of the effective integration of the language model and the acoustic modeling has not yet been achieved due to the interaction between the two knowledge sources preventing estimation of the proper least upper bound of the theory likelihoods. Once this problem is overcome, the stack decoder should become a practical structure for speech recognition.</Paragraph>
  </Section>
class="xml-element"></Paper>