<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1620">
  <Title>An Evolutionary Approach to Referring Expression Generation and Aggregation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In this paper we present a first approach to the idea of using Natural Language Generation (NLG) and Evolutionary Algorithms (EAs) together.</Paragraph>
    <Paragraph position="1"> To test the feasibility of our idea, we decided to select only some particular features of the text on which to put it to the test. Given the complexity of all the changes that are possible to a text, at the levels of syntax, semantics, discourse structure and pragmatics, it seemed impractical to tackle them all at once. For the purpose of illustration, we decided that the problems of the referring expressions and the aggregation were the most suitable to be solved using EAs. Referring Expression Generation involves deciding how each element ocurring in the input is described in the output text. Aggregation involves deciding how compact the presentation of information should be in a given text. It operates at several linguistic levels, but we only consider it here with respect to concepts and their attributes. For instance, the system must decide between generating &amp;quot;The princess is blonde. She sleeps.&amp;quot; and generating &amp;quot;The blonde princess sleeps.&amp;quot;. Aggregation is generally desirable, but may result in adjective-heavy texts when the information to impart becomes dense in terms of attributes, as in &amp;quot;The pretty blonde princess lived in a strong fancy castle with her stern rich parents.&amp;quot;. It is necessary to find the balance between the use of compound or single sentences, or in the case of the modifiers of a concept between the description of the attributes of the concept using only a phrase or various.</Paragraph>
    <Paragraph position="2"> We analysed the features of a human generated text from the point of view of the referring expressions, and we found five different features of simple texts that might be susceptible of easy treatment by means of evolutionary techniques. They are described below.</Paragraph>
    <Paragraph position="3"> Correct Referent.</Paragraph>
    <Paragraph position="4"> When writing a text, we cannot use a pronoun for something that we have not mentioned before, or readers would get confused. An example could be: She lived in a castle. A princess was the daughter of parents.</Paragraph>
    <Paragraph position="5"> In addition, if the full noun reference and the pronoun are far, the reader can also get confused and be unable to link the two occurrences of the same concept, as we can see in the following text: A princess lived in a castle. She was the daughter of parents. She loved a knight. She was pretty.</Paragraph>
    <Paragraph position="6"> She was blonde. It had towers. It was strong. They lived in it.</Paragraph>
    <Paragraph position="7"> Redundant Attributes.</Paragraph>
    <Paragraph position="8"> When describing a concept in an &amp;quot;X is Y&amp;quot; sentence, people do not use the attribute they are going to describe in the reference to the concept. Sentences such as the one below are incorrect: The blonde princess was blonde.</Paragraph>
    <Paragraph position="9"> Reference Repetition.</Paragraph>
    <Paragraph position="10"> Using always the same reference together with the same set of attributes results in repetitive text. For example, it is acceptable to use &amp;quot;the princess&amp;quot; every time we refer to the princess character, but it would be striking to use always &amp;quot;the pretty princess&amp;quot;, as in this example: A pretty princess lived in a castle.</Paragraph>
    <Paragraph position="11"> The pretty princess was the daughter of parents. The pretty princess loved a knight.</Paragraph>
    <Paragraph position="12"> The pretty princess was blonde.</Paragraph>
    <Paragraph position="13"> To avoid that, repetitive use of references is penalized.</Paragraph>
    <Paragraph position="14"> Coherence.</Paragraph>
    <Paragraph position="15"> If we use different subsets of attributes in different references to the same concept, the reader may mistakenly assume that we are referring to different concepts. For example, if we use &amp;quot;the pretty princess&amp;quot; and &amp;quot;the blonde princess&amp;quot; in different places, and we have not specified that the princess is both pretty and blonde, it could seem that there are two princess, a pretty one and a blonde one: A princess lived in a castle. The pretty princess was the daughter of parents. The blonde princess loved a knight.</Paragraph>
    <Paragraph position="16"> Overlooked Information.</Paragraph>
    <Paragraph position="17"> When processing the conceptual representation of a given input, some information about a concept may disappear from the final text. This should be avoided.</Paragraph>
    <Paragraph position="18"> This paper describe an evolutionary solution that guarantees the satisfaction of these restrictions in the conceptual rendition of a given input by means of shallow techniques that rely on very little knowledge about the domain and no reasoning or common sense capabilities.</Paragraph>
  </Section>
class="xml-element"></Paper>