<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-3042">
  <Title>Bi-directional LR Parsing fi'om an Anchor Word for Speech Recognition</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Parsing a word lattice produced by a speech recognition module requires much more search them conventional semence parsing, and thmvt'ore an extremely efficient par:dng algorithm is needed. A word lattice is a set of words hypothesized t,~y a speech recognition system from an utterance. A typical word lattice consists of 30 - 200 words for a 10 word utterance, and each word has a score indicating probability of its having been actually uttered.</Paragraph>
    <Paragraph position="1"> Not only are there many junk words which were never utteced, some actually uttered words may not be present in the lattice (missing words).</Paragraph>
    <Paragraph position="2"> A, island growing parsing in A'IN mechanism presented the serious maintenance and practical problems \[10\]. The first promising allempt to pmse an incomplete word lattice was made by Itayes et al. \[2\], using semantic caseframes. This attempt revealed that, while the semantic caseframes can provide a reasonable degree of robustness, a very efficient algori\[hm is required to be practical. Good efforts were made by Poesio et al. \[4\] and Giachin et al. \[1\] to make the semm~tic caseframe approach more efficient and robust. Meanwhile, Tomita modified the generalized LR parsing algorithm (GLR)\[8\] to handle word lattices \[91\]. The GLR algorithm is a very efficient, table-driven, non-deterministic context-free parsing algorithm, and it has been applied in speech recognition projects with fl~rther modification of the algorithm to handle missing words \[5\].</Paragraph>
    <Paragraph position="3"> It requires heavy search, however, especially when a word is missed in the beginning part of the utterance, since the parser guesses missing words only from its left context.</Paragraph>
    <Paragraph position="4"> Thus, the strict left-to-right-heSS sometimes suffers inefficieucy, and it is desired to parse occasionally backwards from an acoustically reliable word called an anchor word \[10\], Bidirectionality ,also plays an imporlant role in Head-Driven parsing and a method of bi-directional parsing was presented by Satta et al \[7\].</Paragraph>
    <Paragraph position="5"> This paper describes a technique, called bi-directio~ml GLR parsing, to Imrse a word lattice occasionally backwards without loss of the ruble-driven efficiency. A reverse LR table is constructed as well as a standard LR table. Section 2 reviews the generalized LR parsing algorithm. Section 3 then describes how to consU'uct reverse LR tables and how to use them in word lattice parsing. Section 4 discusses the robustness of bi-directional GLR parsing, and finally concluding remarks are made in Section 5.</Paragraph>
  </Section>
class="xml-element"></Paper>