<?xml version="1.0" standalone="yes"?>
<Paper uid="C96-1033">
  <Title>FeasPar - A Feature Structure Parser Learning to Parse Spoken Language</Title>
  <Section position="3" start_page="0" end_page="188" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> When building a speech parsing component for small domains, an important goal is to get good performance. If low hand labor is involved, then it's even better.</Paragraph>
    <Paragraph position="1"> Unification based formalisms, e.g.(Gazdar et al., 1985; Kaplan and Bresnan, 1982; Pollard and Sag, \]987), have been very successful for analyzing written language, because they have provided parses with rich and detailed linguistic information. However, these approaches have two major drawbacks: first, they require hand-designed symbolic knowledge like lexica and grammar rules, and second, this knowledge is too rigid, causing problems with ungranlmaticality and other deviations from linguistic rules. These deviations are manageable and low in number, when analyzing written language, but not for spoken language.</Paragraph>
    <Paragraph position="2"> The latter also contains spontaneous effects and speech recognition errors. (On the other hand, the good thing is that spoken language tend to contain less complex structures than written language.) Several methods have been suggested compensate for these speech related problems: e.g. score and penalties, probabilistic rules, and skipping words (Dowding et al., 1993; Seneff, 1992; Lavie and Tomita, 1993; Issar and Ward, 1993).</Paragraph>
    <Paragraph position="3"> A small community have experimented with either purely statistical approaches(Brown et al., 1990; Schiitze, 1993) or connectionist based approaches (Berg, 1991; Miikkulainen and Dyer, 1991; Jain, 1991; Wermter and Weber, 1994).</Paragraph>
    <Paragraph position="4"> The main problem when using statistical approaches for spoken language processing, is the large amounts of data required to train these models. All connectionist approaches to our knowledge, have suffered from one or more of the following problems: One, parses contains none or too few linguistic attributes to be used in translation or understanding, and/or it is not shown how to use their parse formalism in a total NLP system.</Paragraph>
    <Paragraph position="5"> Two, no clear and quantitative statement about overall performance is made. Three, the approach has not been evaluated with real world data, but with highly regular sentences. Four, millions of training sentences are required.</Paragraph>
    <Paragraph position="6"> In this paper, we present a parser that produces complex feature structures, as known from e.g.</Paragraph>
    <Paragraph position="7"> GPSG(Gazdar et al., 1985). This parser requires only minor hand labeling, and learns the parsing task itself. It generalizes well, and is robust towards spontaneous effects and speech recognition errors.</Paragraph>
    <Paragraph position="8"> The parser is trained and evaluated with the Spontaneous Scheduling Task, which is a negotiation situation, in which two subjects have to decide on time and place for a meeting. The subjects' calendars have conflicts, so that a few sug- null gestions have to go back and tbrth before finding a time slot suitable for both. The data sets are real-world data, containing spontaneous speech effects. 3?he training set consists of 560 sentences, the deveJopment test set of 65 sentences, and the unseen evaluation set of 120 sentences. For clarity, tile examl)le sentences in this paper are among the simpler in the training set. The parser is trained with transcribed data only, but evaluated with transcribed and speech data (including speech recognition errors). The parser produces feature structures, holding semantic information. Feature structures are used as interlingua in the JANUS speech-to-speech translation system(Woszczyna el; al., 1994). Within our research team, the design of the interlingua ILT was determined by the needs of uniticatkm based parser and generator writers. Consequently, the ILT design was ,lot tuned towards connectkmist systeins. On the contrary, our parser must learn the form of tile output provided by a unitication based parser.</Paragraph>
    <Paragraph position="9"> This paper is organized as follows: First, a short tutorial on feature structures, and how to build them. Second, we describe the parser architec~ ture and how it works. Third, we describe the lexicon. Fourth, we describe the tmrser's neural aspects. Fifth, a search algorithm is motivated.</Paragraph>
    <Paragraph position="10"> Then results and conclusion follow.</Paragraph>
  </Section>
class="xml-element"></Paper>