<?xml version="1.0" standalone="yes"?>
<Paper uid="H86-1022">
  <Title>A LOGICAL-FORM AND KNOWLEDGE-BASE DESIGN FOR NATURAL LANGUAGE GENERATION 12 /</Title>
  <Section position="3" start_page="0" end_page="231" type="intro">
    <SectionTitle>
1. INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> we have as a general goal the development of natural language generation capabilities. Independent software systems will state demands to the generation facility in a mutually convenient form. The generator will use those demands to create natural language sentences. Instead of merging generation with other functions of the overall computer system, this design allows for reuse of the generator in other systems, specialized processing of linguistic information, and modular development.</Paragraph>
    <Paragraph position="1"> Our design requires a notation to represent expressive demands. The notation should be of general applicability. For example, a good notation ought to be acceptable as the output of a natural language parser. The notation should have a well-defined semantics. In addition, the generator has to have some way of interpreting the demands. This interpretation has to be efficient.</Paragraph>
    <Paragraph position="2"> In our research, we have used formal logic as a demand language. Network knowledge-bases are used to define the domain of discourse in order to help the generator interpret the logical forms. And a restricted, hybrid knowledge representation is utilized to analyze demands for expression using the knowledge base.</Paragraph>
    <Paragraph position="3"> Arguments for these decisions include the following: Formal logic is a well established means of expressing information with a well-defined semantics. Furthermore, it is commonly used in natural language analyzers and discourse processors, as well as other AI systems. Network knowledge-base notations have been shown to be effective and efficient in language processing. Work on network representations has shown that they too can be given formal semantics \[Schmolze and Lipkis 83\]. Finally, recent work on hybrid knowledge representation systems has shown how to combine the reasoning of logic and network systems \[Brachman 85\]. Restricted-reasoning hybrid systems have shown this reasoning can be done efficiently.</Paragraph>
    <Paragraph position="4"> On our project, we have:  LF and KB Design for Generation Others have attempted to design an interface between a linguistic generation engine and an associated .software system using an appropriate information representation \[Goldman 75, Appelt 83, Hovy 85, Kukich 85, Jacobs 85, McKeown 85\]. Still others have depended oninformation demand representations with similar welldefined semantics and expressive power,e.g., \[Shapiro 79\]. However, he produces a logician's reading of expressions rather than colloquial English. For example, the popular song &amp;quot;Every man loves a woman.&amp;quot;, might be rendered &amp;quot;For all men there exist a woman that they love.&amp;quot;. The generation component of HAM-ANS \[Hoeppner et al. 83} and one effort of McDonald's \[McDonald 83\] are probably closest to our design. HAM-ANS also uses a logical language (the same one used for representing the analyzed input), has an extensive network domain model, and has a separable linguistic engine (although not as broad in coverage as Nigel). However, the interface language is close to surface linguistic representation, e.g., there are particular expressions for tense and voice. So while it is easier to generate sentences from such structures, it is correspondingly harder for software systems to produce demands for expressions without having access to significant amounts of linguistic knowledge.</Paragraph>
    <Paragraph position="5"> McDonald accepts statements in the first order predicate calculus, processes them with a grammar, and outputs excellent English forms. It is hard to evaluate the coverage of McDonald's grammar, however, the program does depend on extensive procedurally-encoded domain-dependent lexical entries. Our domain dependancies are limited to the correct placement of concepts in the NIKL hierarchy and the association of lexical entries with the concepts. These lexical entries are only characterized by syntactic features.</Paragraph>
    <Paragraph position="6"> In Section 2, we present the component technologies we have applied. Section 3 presents the method by which they are combined. Section 4 presents several examples of their use. We conclude with a section describing the open problems identified by our experiences and our plans for future work.</Paragraph>
  </Section>
class="xml-element"></Paper>