<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2014">
  <Title>The Effectiveness of Corpus-Induced Dependency Grammars for Post-processing Speech*</Title>
  <Section position="3" start_page="0" end_page="102" type="intro">
    <SectionTitle>
2 Our System
</SectionTitle>
    <Paragraph position="0"> We have developed the prototype spoken language system depicted in Figure 1 that integrates a speech recognition component based on HMMs with a powerful grammar model based on Constraint Dependency Grammar (CDG). The speech recognizer is implemented as a multiple-mixture triphone HMM with a simple integrated word co-occurrence gram- null mar (Ent, 1997; Young et al., 1997). Mel-scale cepstral coefficients, energy, and each of their their first and second order differences are used as the underlying feature vector for each speech frame. Model training is done using standard Baum-Welch Maximum Likelihood parameter re-estimation on diagonal covariance Gaussian Mixture Model (GMM) feature distributions. The speech recognizer employs a token-passing version of the Viterbi algorithm (Young et al., 1989) and pruning settings to produce a pruned recognition lattice. This pruned lattice contains the most likely alternative sentences that account for the sounds present in an utterance as well as their probabilities. Without any loss of information, this lattice is then compressed into a word graph (Harper et al., 1999b; Johnson and Harper, 1999), which acts as the interface between the recognizer and the CDG parser. The word graph algorithm begins with the recognition lattice and eliminates identical subgraphs by iteratively combining word nodes that have exactly the same preceding or following nodes (as well as edge probabilities), pushing excess probability to adjacent nodes whenever possible. The resulting word graph represents all possible word-level paths without eliminating or adding any paths or modifying their probabilities.</Paragraph>
    <Paragraph position="1"> Word graphs increase the bandwidth of useful acoustic information passed from the HMM to the CDG parser compared to most current speech recognition systems.</Paragraph>
    <Paragraph position="2"> The CDG parser parses the word graph to identify the best sentence consistent with both the acoustics of the utterance and its own additional knowledge.</Paragraph>
    <Paragraph position="3"> The loose coupling of the parser with the HMM allows us to construct a more powerful combined system without increasing the amount of training data for the HMM or the computational complexity of either of the component modules. Our NLP component is implemented using a CDG parser (Harper and Helzerman, 1995; Maruyama, 1990a; Maruyama, 1990b) because of its power and flexibility, in particular: * It supports the use of syntactic, semantic, and domain-specific knowledge in a uniform framework. null * Our CDG parser supports efficient simultaneous parsing of alternative sentence hypotheses in a word graph (Harper and Helzerman, 1995; Helzerman and Harper, 1996).</Paragraph>
    <Paragraph position="4"> * Because CDG is a dependency grammar, it can better model free-order languages. Hence, CDG can be used in processing a wider variety of human languages than other grammar paradigms.</Paragraph>
    <Paragraph position="5"> * It is capable of representing and using context-dependent information unlike traditional grammar approaches, thus providing a finer degree of control over the syntactic analysis of a sentence.</Paragraph>
    <Paragraph position="6"> * A CDG can be extracted directly from sentences annotated with dependency information (i.e., feature and syntactic relationships).</Paragraph>
    <Paragraph position="7"> We hypothesize that the accuracy of the combined HMM/CDG system should benefit from the ability to create a grammar that covers the domain as precisely as possible and that does not consider sentences that would not make sense given the domain. A corpus-based grammar is likely to have this degree of control. In the next section we describe how we construct a CDG from corpora.</Paragraph>
    <Paragraph position="8">  ken language system.</Paragraph>
  </Section>
class="xml-element"></Paper>