<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1405">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Individuality and Alignment in Generated Dialogues</Title>
  <Section position="4" start_page="0" end_page="25" type="intro">
    <SectionTitle>
2 Background
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="25" type="sub_section">
      <SectionTitle>
2.1 Personality and Language
</SectionTitle>
      <Paragraph position="0"> Current work on personality traits is dominated by Costa and McCrae's five-factor model (Costa and McCrae, 1992). The five factors, or dimensions, are: Extraversion; Neuroticism; Openness; Agreeableness; and Conscientiousness (Matthews et al., 2003). It has been shown that scores on these dimensions correlate with some aspects of language use (Scherer, 1979; Dewaele and Furnham, 1999).</Paragraph>
      <Paragraph position="1"> In studies of text, the focus has been on lexical choice, and Pennebaker and colleagues have analysed relative frequencies of use of word-stems in a dictionary structured into semantic and syntactic categories (Pennebaker et al., 2001). Amongst other results, they have shown that High Extraverts  use: more social process talk, positive emotion words and inclusives; and fewer negations, tentative words, exclusives, causation words, negative emotion words, and articles (Pennebaker and King, 1999; Pennebaker et al., 2002).</Paragraph>
      <Paragraph position="2"> Computational linguistic exploitation of such empirically-derived features has been limited. On the one hand, in generation, there has been work on personality-based generation. For instance, in developing embodied conversational agents, researchers have designed agents or teams of agents with distinguishable linguistic personalities (Ball and Breese, 2000; Rist et al., 2003; Piwek and van Deemter, 2003; Gebhard, 2005). However, the linguistic behaviour is usually informed by rules based on personality stereotypes, rather than on language statistics themselves. On the other hand, in interpretation, more empirical work has recently been carried out, to enable text classification. Argamon et al. (2005) attempted to classify authors as High or Low Extravert and High or Low Neurotic, using Pennebaker and King's (1999) data. They report classification accuracies of around 58% (with a 50% baseline). Oberlander and Nowson (2006) undertake a comparable task, using weblog data. They report classification accuracies of roughly 85% (Neuroticism) and 94% (Extraversion), and comparable figures for Agreeableness and Conscientiousness. Such studies can provide ordered lists of linguistic features which are useful for distinguishing language producers, and we will return to this, below.</Paragraph>
    </Section>
    <Section position="2" start_page="25" end_page="25" type="sub_section">
      <SectionTitle>
2.2 Alignment and Language
</SectionTitle>
      <Paragraph position="0"> People converge with their interlocutors in linguistic choices at a number of levels (Pickering and Garrod, 2004). The phenomena can be seen in both social and cognitive terms. On the social side, co-operative processes such as audience design are usually considered to be conscious, at least in part (Bell, 1984). But on the cognitive side, coordinative processes such as alignment are usually considered to be largely automatic (Garrod and Doherty, 1994). Alignment can be probed by psycholinguistic tests for interpersonal priming, establishing the extent to which participants are more likely to use a lexical item or syntactic construction after hearing their conversational partner use it. Syntactic priming experiments involve constructions such as passives, and ditransitives (Pickering and Branigan, 1998).</Paragraph>
      <Paragraph position="1"> It is possible that some people are stronger aligners than others. Gill et al. (2004) probed syntactic priming for passives, and investigated whether levels of Extraversion or Neuroticism would affect the strength of priming effects. It was found that Extraversion has no effect, but that Neuroticism has a non-linear effect: both High and Low levels of Neuroticism led to weaker priming; Mid levels led to significantly stronger priming.</Paragraph>
      <Paragraph position="2"> Given this, if a generation system is going to simulate alignment, it is probably worth designing it so that it can simulate agents with differing propensities to align.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>