<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1017">
  <Title>A Constructive View of GPSG or How to Make It Work</Title>
  <Section position="2" start_page="0" end_page="77" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Any attempt to build a multi-lingual MT system as in EUROTRA \[King, Perschke 1987\] must provide for massive modularization in order to avoid developing 9 parsers, 9 generators and 72 transfer components for the 9 languages involved, not to mention the different but redundant formulations of linguistic knowledge embodied in them. The most obvious approach consists in developing one single parser, one single generator, and one single transfer component, the first two being capable of dealing with grammars for different languages and the latter with transfer 1 This work has been developed in the project KFr-FAST (KIT = Kilnstliche Intelligenz und Textverstehen (Artificial Intelligence and Text Understanding); FAST = Functor Argument Structure for Translation), which constitutes the Berlin component of the complementary research project of EuroWa-D. It receives grants by the Federal Minister for Research anti Technology under contract 1013211.</Paragraph>
    <Paragraph position="1"> rules for different pairs of languages. Moreover, an MT system must be based on a linguistically justified theory of grammar.</Paragraph>
    <Paragraph position="2"> This theory has to be implemented in the system, where it determines the construction of a syntactic representation of a sentence during the parsing of some input string as well as during the generation based on the output of the transfer component.</Paragraph>
    <Paragraph position="3"> The theory of GPSG (see \[Gazdar et al. 1985\], henceforth: \[GKPS\]) has been tested for its usefulness for MT \[ttauenschild/Busemann 1988\]. It offers the high degree of modul,'u'ity that is required. For instance, an implementation of the GPSG formalism would be able to run with different grammars, and linguistic generalizations would either evolve from the formalism (in the case of universals), or be expressible within the grammars (in the case of language-specific generalizations). We shall distinguish between the formalism and the grammar in the following way; the formalism consists of the Feature Instantiation Principles (FIPs), the formal definition of syntactic features, categories, Feature Co-occurrence Restrictions (FCRs), Immediate Dominance (ID) rules, Linear Precedence (LP) statements, admissible trees, etc. The grammars consist of actual sets of ID rules, LP statements, FCRs, and the lexicon.</Paragraph>
    <Paragraph position="4"> However, a closer look at the axiomatic way GPSG has been defined reveals severe problems for an implementation of GPSG. In the next section we shall outline these problems, and in section 3 present our change in perspective towards a GPSG formalism that overcomes these problems. Some consequences of this are discussed in the last section.</Paragraph>
    <Paragraph position="5"> The rest of the paper concentrates on GPSG and its use for processing of representations of natural language sentences.</Paragraph>
    <Paragraph position="6"> Nothing can be said here about the necessity of including textual knowledge for translation or about the transfer step itself (but cf. \[Hauenschild 1986\]).</Paragraph>
  </Section>
class="xml-element"></Paper>