<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1422">
  <Title>Enriching Partially-Specified Representations for Text Realization Using an Attribute Grammar *</Title>
  <Section position="2" start_page="0" end_page="163" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Typically, a text realization system requires a great deal of syntactic information from an application in order to generate a high quality text; however, an application might not have this information (unless it has been built with text generation in mind).</Paragraph>
    <Paragraph position="1"> This problem has been referred to as the Generation Gap (Meteer, 1990). Meteer first identified the generation gap problem as arising at the text planning stage. A text planner must decide what content needs to be expressed and creates a corresponding text plan for generating it. A sentence planner is then used to select an appropriate syntactic strucdegThis work was supported by a gift from Intel Corporation; and by the National Science Foundation, under grants IRI9701617 and IRI-9523666.</Paragraph>
    <Paragraph position="2"> ture for a given plan. Typically, neither a text planner nor a sentence planner is concerned with fine-grained syntactic issues, such as whether the subject of the sentence is a singular or plural noun. Thus, it becomes the responsibility of a text realizer to infer the missing information and to generate the best possible text from a given input.</Paragraph>
    <Paragraph position="3"> Most generation systems (such as FUF/SURGE (Elhadad, 1992), Penman (Mann, 1983), Real-Pro (Lavoie and Rainbow, 1997), TG/2 (Busemann, 1996), and YAG (Channarukul, 1999; McRoy et al., 1999)) alleviate this problem by using defaulting, in which a grammar writer specifies a default for each syntactic constraint. This approach is inflexible and prone to errors, because there might not be one default that suits all applications or situations.</Paragraph>
    <Paragraph position="4"> Another approach that has been proposed is to fill in the missing information on the basis of word co-occurrence data collected from a large corpus of text (see Nitrogen (Knight and Hatzivassiloglou, 1995)).</Paragraph>
    <Paragraph position="5"> However, statistical approaches have difficulty when there are long-distance dependencies among constituents in a text.</Paragraph>
    <Paragraph position="6"> In this paper, we present a new approach to resolving the so-called generation gap that uses an Attribute Grammar (Knuth, 1968) to enrich partially-specified inputs to a realization system to produce high quality texts. Attribute Grammars are a declarative formalism for defining rules for attribute propagation (see Section 3). They have been used primarily for specifying the .semantics of programruing languages, although a few researchers have also used them to drive a text generator (see (Levison and Lessard, 1990), for exaanple). The main advantage of our approach is that it allows a generator to enjoy the computational efficiency of a template-based realization system, while reducing the linguistic burden on an application and increasing the quality of the generated texts.</Paragraph>
    <Paragraph position="7"> Our work differs from previous uses of attribute grammars in natural language generation, which are similar to Levison and Lessard (Levison and Lessard, 1990)in that they apply attribute grammars directly to text realization. For example, Lev- null the Sentence &amp;quot;Jack and I want his sister's dog to swim.&amp;quot; ison and Lessard extend a context-free grammar with attributes and semantic rules similar to classical attribute grammars presented by Knuth (Knuth, 1968). Attributes in their system assist the realization by propagating information down a tree that specifies the complete syntactic structure of the output text. By contrast, our work employs attribute grammars, not to realize a text, but to perform a generation gap analysis prior to actual realization.</Paragraph>
    <Paragraph position="8"> We use both inherited and synthesized attributes (i.e., propagating information both down and up a tree) to share information and to determine appropriate values for any missing features.</Paragraph>
  </Section>
class="xml-element"></Paper>