<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-1026">
  <Title>TINA: A PROBABILISTIC SYNTACTIC PARSER FOR SPEECH UNDERSTANDING SYSTEMS*</Title>
  <Section position="3" start_page="0" end_page="168" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> Most syntactic parsers have been designed with the assumption that the input word stream is deterministic: i.e., at any given point in the parse tree it is known with certainty what the next word is. As a consequence, these parsers generally cannot be used effectively, if at all, to provide syntax-directed constraint in the speech recognition component of a spoken language system. In a fully integrated system, the recognizer component should only be allowed to propose partial word sequences that the natural language component can interpret. Any word sequences that are syntactically or semantically anomalous should probably be pruned prior to the acoustic match, rather than examined for approval in a verification mode. To operate in such a fully integrated mode, a parser should have the capability of considering a multitude of hypotheses simultaneously. The control strategy should have a sense of which of these hypotheses, considering both linguistic and acoustic evidence, is most likely to be correct at any given instant in time, and to pursue that hypothesis only incrementally before reexamining the evidence. The linguistic evidence should include probability assignments on proposed hypotheses; otherwise the perplexity of the task is much too high for practical recognition applications.</Paragraph>
    <Paragraph position="1"> This paper describes a natural language system, TINA, which addresses many of these issues. The grammar is constructed by converting a set of context-free rewrite rules to a form that merges common elements on the right-hand side (RHS) of all rules sharing the same left-hand side (LHS). Elements on the LHS become parent nodes in a family tree. Through example sentences, they acquire knowledge of who their children are and how they can interconnect. Such a transformation permits considerable structure sharing among the rules, as is done in typical shift-reduce parsers \[3\]. Probabilities are established on arcs connecting pairs of right siblings rather than on rule productions. We believe this is a more reasonable *This research was supported by DARPA under Contract N00039-85-C-0254, monitored through Naval Electronic Systems Command.</Paragraph>
    <Paragraph position="2">  way to use probabilities in a grammar. Context-dependent constraints to deal with agreement and gaps are realized by passing features and flags among immediate relatives.</Paragraph>
  </Section>
class="xml-element"></Paper>