<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1072">
  <Title>The Test Results</Title>
  <Section position="3" start_page="0" end_page="366" type="intro">
    <SectionTitle>
LANGUAGE CONSTRAINTS
DURING RECOGNITION
</SectionTitle>
    <Paragraph position="0"> In order to obtain adequate recognition results for continuous speech recognition, it is imperative to provide some sort of language constraints. The usual approach is to adopt simple but efficient word-pair or bigram &amp;quot;language models,&amp;quot; which specify the set of words that can follow a given word.</Paragraph>
    <Paragraph position="1"> Such models have the advantage of being automatically derivable from training data and computationally efficient. However, they lose any non-local language constraints and, of course, provide no linguistically relevant structural description. Furthermore, it is difficult, even when steps are taken to generalize words to their semantic category (i.e. &amp;quot;Boston&amp;quot; all city names), to assure sufficient coverage in an independent test set. Given adequate training data, the simple word-pair or bigram language model will overgenerate, since it fails to take larger context into account. Thus the system is allowed to recognize many sentences that are ungrammatical or incoherent or inappropriate in the overall context.</Paragraph>
    <Paragraph position="2"> The obvious solution is to bring linguistic knowledge to bear. One way is to take the best acoustic candidate and use a flexible, semantically-based phrase-spotting system to assign a meaning to the sequence of words \[8\]. This provides a robust interface which can ignore many recognition errors and abandons the notion of a linguistically well-formed over-all sentence. It almost always produces some interpretation.</Paragraph>
    <Paragraph position="3"> However, since it adds no real linguistic constraints, it may produce many false positives (misinterpretation of the input ) .</Paragraph>
    <Paragraph position="4">  A second possiblity which has been explored at some sites \[1\] is to have the recognizer produce a word lattice, with (acoustic) transition probabilities between words. The language system can then search this lattice for the best candidate.</Paragraph>
    <Paragraph position="5"> Another approach, which is the baseline for these experiments, uses an N-best interface between the recognizer and the language understanding system. In this interface, the recognizer produces sentence hypotheses in decreasing order of acoustic score. The role of the language understanding system is to filter these hypotheses, choosing the first one that can be fully processed. Finally, the approach that we explore here combines a score provided by the parser (e.g, on the basis of a probability assignment), with the acoustic score, to provide a &amp;quot;best&amp;quot; answer. We will report here on the results of several experiments combining parse probabilities and acoustic score.</Paragraph>
  </Section>
class="xml-element"></Paper>