<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0319">
  <Title>Has a Consensus NL Generation Architecture Appeared, and is it Psycholinguistically Plausible?</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In this paper I survey some recently-developed NL generation systems that (a) cover the complete generation process and (b) are designed to be used by application programs, as well as (or even instead of) making some theoretical point. I claim that despite their widely differing theoretical backgrounds, the surveyed systems are similar in terms of the modules they divide the generation process into, the way the modules interact with each other, and (at least in some cases) the kinds of computations each individual module performs. In other words, despite different theoretical claims, there is a remarkable level of similarity in how these systems 'really work'; that is, a de facto 'consensus architecture' seems to be emerging for how applied NLG systems should generate text. The existence of such agreement among the surveyed systems is especially *Most of this work was done while the author was at the University of Edinburgh, Department of Artiticia.l Intelligence. The Edinburgh work was supported by SERC grant GR/F/36750.</Paragraph>
    <Paragraph position="1"> surprising because in some cases the theoretical backgrounds of the systems examined argue against some aspects of the consensus architecture.</Paragraph>
    <Paragraph position="2"> I also compare the consensus architecture to psycholinguistic knowledge about language generation in human speakers. Such a comparison is often difficult to make, because of the many gaps in our current knowledge about how humans speak. Nevertheless, I argue that as far as such a comparison can be made, the specific design decisions embodied in the consensus architecture seem to often be more or less in accord with current knowledge of human language generation. This is again perhaps somewhat surprising, since psycholinguistic plausibility was not in general a goal of the developers of the examined systems. Perhaps (being very speculative) this indicates that there is some connection between the engineering considerations that underlie the design decisions made in the consensus architecture, and the maximizeperformance-in-the-real-world criteria that drove the evolutionary processes that created the human language processor. If (a big if!) there is some truth to this hypothesis, then studying the engineering issues involved in building applied systems may lead to insights about the way the human language system works.</Paragraph>
  </Section>
class="xml-element"></Paper>