<?xml version="1.0" standalone="yes"?>
<Paper uid="A88-1006">
  <Title>FROM WATER TO WINE: GENERATING NATURAL LANGUAGE TEXT FROM TODAY'S APPLICATIONS PROGRAMS 1</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> Work in artificial intelligence has two goals: on the one hand to do concrete work that can be used in actual systems today, on the other to establish strong theoretical foundations that will allow us to build more sophisticated systems tomorrow.</Paragraph>
    <Paragraph position="1"> Unfortunately, since the field is so young and so few problems are well understood, these two goals are often at odds.</Paragraph>
    <Paragraph position="2"> Natural language generation is no exception. The years of research in linguistics have made problems in syntax comparatively well understood.</Paragraph>
    <Paragraph position="3"> Nevertheless, we should not restrict ourselves to just generating single, isolated sentences until the problems of lexical semantics, discourse structure, and conceptual modeling are understood as well. We must find ways to facilitate both efforts, modularizing our systems so that the parts that handle well understood processes need not be compromised to accomodate weaknesses in other parts of the system. This paper is on how to support such modularity in a natural language generator.</Paragraph>
    <Paragraph position="4"> 1 This work was supported in part by DARPA contracts N00014-87-K0238 at the Univerisity of Massachusetts and DAAA 15-87-C0006 CDRLA002 at BBN Laboratories, and by the Rome Air Development Center contract number AF30602- 81-C-0169, task number I74398 at the University of Massachusetts.</Paragraph>
    <Paragraph position="5"> In the present case, the well understood process is linguistic realization, and the weaknesses are in the conceptual models and representations of the programs underlying the generator. To bridge this gap, we present a specification language, to be used as input to the linguistic realization component Mumble-86. 2 This language provides the designer of a planning component with a vocabulary of linguistic resources (i.e. words, phrases, syntactic constructions) and a straightforward means of directing their composition. The specification language facilitates interfacing Mumble to a wide range of underlying programs and planners. For simple programs not built with language in mind, we show a straightforward means of using predefined templates to map underlying objects to complex linguistic structures. For systems with more sophistication in text planning, we show how the compositionality and flexibility of the specification language can be used to make their task easier. What is template driven at one end of the range can be built compositionally at the other; what is stipulated at one end can be reasoned about at the other.</Paragraph>
  </Section>
class="xml-element"></Paper>