<?xml version="1.0" standalone="yes"?>
<Paper uid="P94-1016">
  <Title>INTERLEAVING SYNTAX AND SEMANTICS IN AN EFFICIENT BOTTOM-UP PARSER*</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> The parsing problem is typically framed as a recognition problem: Given a grammar and a word string, determine if the word string is a member of the language described by the grammar. For some applications, notably robust natural-language processing and spoken-language understanding, this is insufficient, since many utterances will not be accepted by the grammar, because of nonstandard language, inadequate grammatical coverage, or errors made in speech recognition. In these cases, it is still desirable to determine what well-formed phrases occurred {n the word string, even when the entire string is not recognized. The goal of the parser described here is to construct a chart, as efficiently as possible, that contains all the syntactically well-formed semantically meaningful phrases *This research was supported by the Advanced Research Projects Agency under Contract ONR N0001490-C-0085 with the Office of Naval Research. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the Advanced Research Projects Agency of the U.S. Government.</Paragraph>
    <Paragraph position="1"> tCurrent address: CAP GEMINI Innovation, 8690 Rue Thiers, 92513-Boulogne Billancourt, France, andry@capsoget i. fr.</Paragraph>
    <Paragraph position="2"> that occur in the word string.</Paragraph>
    <Paragraph position="3"> The most efficient practical context-free parsers (Earley, 1970; Graham, Harrison, and Ruzzo, 1980) are left-corner parsers, which gain efficiency by their ability to constrain the search to find only phrases that might contribute to a sentence that starts at the left edge of the string being parsed. These strong left-context syntactic constraints can prevent the parser from finding some phrases that are well-formed, however. This is a problem for us that is avoided by bottom-up parsers (Kasami, 1965;Younger, 1967), but at the expense of creating many more edges, which can lead to dramatic increases in parse time.</Paragraph>
    <Paragraph position="4"> Since our goal is to find only the phrases that are semantically meaningful as well as syntactically well-formed, we also need to compute semantic constraints for every syntactic phrase we construct. This requires making finer distinctions than syntax-only parsing, which can introduce additional ambiguity, multiplying the number of distinct phrases found and increasing parse time.</Paragraph>
    <Paragraph position="5"> We describe two special techniques for speeding up bottom-up parsing by reducing local ambiguity without sacrificing completeness. One technique, &amp;quot;limited left-context checking,&amp;quot; reduces local syntactic ambiguity; the other, &amp;quot;deferred sortal-constraint application,&amp;quot; reduces local semantic ambiguity. Both techniques are applied to unification-based grammars. We analyze the performance of these techniques on a 194-utterance subset of the AP~PA ATIS corpus (MADCOW, 1992), using a broad-coverage grammar of English.</Paragraph>
    <Paragraph position="6"> Finally, we present results using the output of the parser to improve the accuracy of a speech recognizer in a way that takes advantage of our ability to find all syntactically well-formed semantically meaningful phrases.</Paragraph>
  </Section>
class="xml-element"></Paper>