<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1423">
  <Title>APPROACHES TO SURFACE REALIZATION WITH HPSG</Title>
  <Section position="3" start_page="0" end_page="219" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In work on natural language generation, tile most influential linguistic framework has l)robably been Systemic Functional Grammar (SFG). However, in other areas of computational linguistics the most widely used grammatical framework appears to be Head-driven Phrase Structure Grarnmar (HPSG). Why is it, then, that using HPSG for generation has been almost as unpopular as using SFG for parsing? Without making any claim that HPSG is better t.han SFG for generation, we will review some plausible approaches to surface realization with HPSG. We will show that there are indeed some fundamental difficulties in using HPSG for generation, but also that there are some solutions to these difficulties.</Paragraph>
    <Paragraph position="1"> The first approach to mention is the radical one of converting HPSG into something else before generation, such as Tree Adjoining Grammar (Kasper et al., 1995)..Though this seems to support the view that HPSG is unsuitable for generation, it is in fact a valuable contribution to work on compiling HPSG grammars for efficient processing, whether for parsing (Torisawa and Tsujii, 1996) or for generation.</Paragraph>
    <Paragraph position="2"> However, we will not be concerned with efficiency, but with more basic problems in the relations between HPSG and generation algorithmS. The question is, can existing algorithms be used with HPSG grammars at all? For clarity, we use the simplest versions of the algorithms, which were originally developed for use with definite clause (DCG) grammars and categorial grammar. For uniformity~ the algorithms are implemented m Prolog and the grammars are implemented in ProFIT (Erbach, 1995). ..</Paragraph>
    <Section position="1" start_page="0" end_page="218" type="sub_section">
      <SectionTitle>
1.1 Generation from what?
</SectionTitle>
      <Paragraph position="0"> A basic problem in using ttPSG for generation is the question &amp;quot;Generation from what?&amp;quot; Various different semantic representations have been used with HPSG granamars, partly due to differences in semai~tic theories and partly due to differences in the requirements of particular applications, such as database interfaces, machine translation, or interactive dialogues.</Paragraph>
      <Paragraph position="1">  The semantic theory which has been particularly associated with HPSG is Situation Semantics. As feature structures became central to linguistic description, and unification became central to linguistic processing, the standard (Pollard and Sag, 1994) semantic representation in HPSG has been a feature structure version of Situation Semantics, and semantic composition has been implemented by unification of the semantic features of the components. So one answer to what generation should start from is to generate from Situation Semantics.</Paragraph>
      <Paragraph position="2"> The distinctive characteristic of HPSG is its emphasis on a head-driven organization of grammar. So it is natural to try using a head-driven generation algorithm, and this is compatible with the feature structure version of Situation Semantics. Head-driven approaches to generation with HPSG are described in detail by Wilcock and Matsumoto (1998). That work is summarised here, in Section 2 where a simple approach runs into fundamental difficulties and in Section 3 where a more sophisticated approach offers a solution. In machine translation, &amp;quot;head-switching&amp;quot; between languages (when the syntactic or semantic head of a source languagestructure does not naturally transfer to the head of a translationally equivalent structure in the target language) means that a strongly head-driven approach to semantics is undesirable. The problem of logical form equivalence is also crucial for generation in machine translation: A flat., list-based semantic representation is therefore more suitable. Minimal Recursion Semantics (Copestake et al., 1997) has been developed specifically to provide such a flat representation for HPSG.</Paragraph>
      <Paragraph position="3"> For generation from fiat lists, we need non=head-driven approaches. In Section 4 we show how an existing bag generation algorithm, developed for use with categorial grammar and indexed logical form, can also be used with HPSG and Minimal Recursion Semantics implemented in ProFIT: In interactive dialogues, generation needs to start from an incomplete bag of semantic terms, and continue incrementally as more terms are added. In Section 5 we describe an approach to incremental generation with HPSG. In contrast to categorial grammar, HPSG has some fundamental difficulties with highly incremental generation. However, we suggest a chart:based solution to the problem.</Paragraph>
      <Paragraph position="4"> In conclusion, some other approaches are briefly mentioned in Section 6. However, before discussing the different approaches to generation, we introduce the ProFIT system used for the implementations.</Paragraph>
    </Section>
    <Section position="2" start_page="218" end_page="219" type="sub_section">
      <SectionTitle>
1.2 HPSG in ProFIT
</SectionTitle>
      <Paragraph position="0"> ProFIT (Erbach. 1995) extends Prolog wit.h typed feature structures. A type hierarchy declaration defines the subtypes and appropriate features of every type. Typed feature terms, which can be mixed with ordinary terms in Prolog procedures, are compiled into normal terms by ProFIT before the procedures are passed to the normal Prolog compiler. An idea of how HPSG can be implemented in ProFIT is shown ill Figure 1, from (Wilcock and Matsumot.o, 1995), where fixrther details are explained. We note here only that the Semantics Principle 'SemP' is defined by a template (:=) which says that the CONTENT of the mother is the same as the CONTENT of the head daughter, and this principle is imposed on all head-nexus-phrases (non-adjunct phrases) by invoking the template by (~:'SemP' within the template for hd_r, exus_ph.</Paragraph>
      <Paragraph position="2"/>
    </Section>
  </Section>
class="xml-element"></Paper>