<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1008">
  <Title>Empirically-based Control of Natural Language Generation</Title>
  <Section position="3" start_page="0" end_page="58" type="intro">
    <SectionTitle>
1 Paiva and Evans (2004) provides an overview of our
</SectionTitle>
    <Paragraph position="0"> framework and detailed comparison with previous approaches to stylistic control (like Hovy (1988), Green and DiMarco (1993) and Langkilde-Geary (2002)). This paper provides a more detailed account of the system and reports additional experimental results. null process. Such architectures have been particularly prominent in the recent development of empirically-based approaches to NLG, where generator outputs can be selected according to application requirements acquired directly from human subjects (e.g. Walker et al. (2002)) or statistically from a corpus (e.g. Langkilde-Geary (2002)).</Paragraph>
    <Paragraph position="1"> However, this approach suffers from a number of drawbacks:  1. It requires generation of all, or at least  many solutions (often hundreds of thousands), expensive both in time and space, and liable to lead to unnecessary interactions with other components (e.g. knowledge bases) in complex systems. Recent advances in the use of packed representations ameliorate some of these issues, but the basic need to compare a large number of solutions in order to rank them remains.</Paragraph>
    <Paragraph position="2">  2. The 'test' component generally does not give fine-grained control -- for example, in a statistically-based system it typically measures how close a text is to some single notion of ideal (actually, statistically average) output.</Paragraph>
    <Paragraph position="3"> 3. Use of an external filter does not combine  well with any control mechanisms within the generator: e.g. controlling combinatorial explosion of modifier attachment or adjective order.</Paragraph>
    <Paragraph position="4"> In this paper we present an empirically-based method for controlling a generator which overcomes these deficiencies. It controls the generator internally, so that it can produce just one (locally) optimal solution; it employs a model of language variation, so that the generator can be controlled within a multidimensional space of possible variants; its view of the generator is completely holistic, so that it can accommodate any other control mechanisms intrinsic to the generation task.</Paragraph>
    <Paragraph position="5">  To illustrate our approach we describe a system for controlling 'style' in the sense of Biber (1988) during the generation of short texts giving instructions about doses of medicine. The paper continues as follows. In SS2 we describe our overall approach. We then present the implemented system (SS3) and report on our experimental evaluation (SS4). We end with a discussion of conclusions and future directions (SS5).</Paragraph>
  </Section>
class="xml-element"></Paper>