<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2311">
  <Title>The Importance of Discourse Context for Statistical Natural Language Generation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The purpose of a natural language generation (NLG) system is to encode semantic content in a linguistic form easily understood by humans in order to communicate it to the user of the system. Ideally, this content should be encoded in strings that are both grammatical and contextually appropriate.</Paragraph>
    <Paragraph position="1"> Human speakers of all natural languages have many ways to encode the same truth-conditional meaning besides a single &amp;quot;canonical&amp;quot; word order, even when encoding one predicate and its arguments as a main clause. Humans choose contextually-appropriate options from these many ways with little conscious effort and with rather effective communicative results. Statistical approaches to natural language generation are based on the assumption that often many of these options will be equally good, e.g. (Bangalore and Rambow, 2000).</Paragraph>
    <Paragraph position="2"> In this paper, we argue that, in fact, not all options are equivalent, based on linguistic data both from English, a language with relatively static word order, and from Finnish, a language with much more flexible word order. We show that a statistical NLG algorithm based only on counts of trees cannot capture the appropriate use of word order. We provide an alternative method which has been implemented elsewhere and show that it dramatically outperforms the statistical approach. Finally, we explain how the alternative method could be used to augment present statistical approaches and draw some lessons for future development of statistical NLG.</Paragraph>
    <Paragraph position="3"> 2 Statistical NLG: a brief summary In recent years, a new approach to NLG has emerged, which hopes to build on the success of the use of probabilistic models in natural language understanding (Langkilde and Knight, 1998; Bangalore and Rambow, 2000; Ratnaparkhi, 2000). Building an NLG system is highly labor-intensive. For the system to be robust, large amounts of world and linguistic knowledge must be handcoded. The goal of statistical approaches is to minimize hand-coding and instead rely upon information automatically extracted from linguistic corpora when selecting a linguistic realization of some conceptual representation.</Paragraph>
    <Paragraph position="4"> The underlying concept of these statistical approaches is that the form generated to express a particular meaning should be selected on the basis of counts of that form (either strings or trees) in a corpus. In other words, in generating a form f to express an input, one wants to maximize the probability of the form, P(f), with respect to some gold-standard corpus, and thus express the input in a way that resembles the realizations in the corpus most closely (Bangalore and Rambow, 2000). Bangalore and Rambow's algorithm for generating a string in the FERGUS system begins with an underspecified conceptual representation which is mapped to a dependency tree with unordered sibling nodes. To convert the dependency tree into a surface form, a syntactic structure is chosen for each node. In FERGUS, this structure is an elementary tree in a tree-adjoining grammar. The choice of a tree is stochastic, based on a tree model derived from 1,000,000 words of the Wall Street Journal. For example, the tree chosen for a verb V will be the most frequently found tree in the corpus headed by V .</Paragraph>
  </Section>
class="xml-element"></Paper>