<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1432">
  <Title>Sentence generation and neural networks</Title>
  <Section position="1" start_page="0" end_page="0" type="abstr">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> In this paper we describe a neural networks approach to generation. The task is to generate sentences with hotel-information from a structured database. The system is inspired by Karen Kukich's ANA, but expands on it by adding generality in the form of language independence in representations and lexical look-up.</Paragraph>
    <Paragraph position="1"> Introduction In the growing field of intelligent communication (web-browsers, dialogue systems, etc.) the need for a flexible generator has become more important (e.g. Hovy &amp; Lin, 1999). NLG is usually seen as a two-stage process where the planning component takes care of the inter-sentential content planning, while the surface realisation component transforms the content representation into a string of words. Interactions between the two components have called for the micro-planning stage to be postulated in the middle, but still the rule-based pipeline architecture has problems with sequential rules and their two-way relations. Statistical approaches have been developed, and seem to provide flexibility to generation tasks.</Paragraph>
    <Paragraph position="2"> The approach taken in this thesis, however, explores generation as .a .classification task whereby the representation that describes the intended meaning of the utterance is ultimately to be classified into an appropriate surface form. Although the task as such is a complex one, the approach allows its decomposition into a series of smaller classification tasks tbrmulated as input-output mappings rather than step-wise rules. One of the goals of the thesis is to study the ways generation could .be broken down into suitable sub-classification tasks so as to enhance flexibility in the generation process in general. Artificial neural networks are a classification technique that is robust and resistant to noisy input, and learns to classify inputs on the basis of training examples, without specific rules that describe how the classification is to be done.</Paragraph>
    <Paragraph position="3"> There is not much research into using ANN's for generation, the main reason being long training times. Two notable exceptions are Kukich (1987) and Ward (1997), both argue in favour of NN's robustness, but at the same time point out problems with scalability. We believe that with improved computer facilities that shorten the training time, this new way of looking at generation as a classification task constitutes an interesting approach to generation. We have chosen Kukich's approach, as our application domain is to generate utterances from structured databases.</Paragraph>
    <Paragraph position="4"> This paper is structured as follows; we first discuss the general model. The second part briefly describes neural networks. We continue with describing a possible implementation of the model, and finally we draw some conclusions and point to future challenges.</Paragraph>
  </Section>
class="xml-element"></Paper>