<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2022">
  <Title>Automatic Recognition of Personality in Conversation</Title>
  <Section position="4" start_page="86" end_page="86" type="evalu">
    <SectionTitle>
3 Results
</SectionTitle>
    <Paragraph position="0"> The features characterize many aspects of language production: utterance types, content and syntax (LIWC), psycholinguistic statistics (MRC), and prosody. To evaluate how each feature set contributes to the final result, we trained models with the full feature set and with each set individually.</Paragraph>
    <Paragraph position="1"> Results are summarized in Table 3. The baseline is a model ranking extracts randomly, producing a ranking error of 0.5 on average. Results are averaged  validation for different feature sets (Type=utterance type, Pros=prosody). Best models are in bold.</Paragraph>
    <Paragraph position="2"> Paired t-tests show that models of extraversion, agreeableness, conscientiousness and intellect using all features are better than the random ordering base-line (two-tailed, p &lt; 0.05)  . Emotional stability is the most difficult trait to model, while agreeableness  We also built models of self-reports of personality, but none of them significantly outperforms the baseline. and conscientiousness produce the best results, with ranking errors of 0.31 and 0.33 respectively. Table 3 shows that LIWC features perform significantly better than the baseline for all dimensions but emotional stability, while emotional stability is best predicted by MRC features. Interestingly, prosodic features are very good predictors of extraversion, with a lower ranking error than the full feature set (0.26), while utterance type features on their own never out-perform the baseline.</Paragraph>
    <Paragraph position="3"> The RankBoost rules indicate the impact of each feature on the recognition of a personality trait by the magnitude of the parameter a associated with that feature. Table 4 shows the rules with the most impact on each best model, with the associated a values. The feature labels are in Table 2. For example, the model of extraversion confirms previous findings by associating this trait with a high speech rate (Rules 1 and 4) and longer conversations (Rule 5). But many new markers emerge: extraverts speak with a high pitch (Rules 2, 6 and 7), while introverts' pitch varies a lot (Rules 15, 18 and 20). Agreeable people use longer words but shorter sentences (Rule 1 and 20), while swear words reduce the agreeableness score (Rules 12, 18 and 19). As expected, conscientious people talk a lot about their job (Rule 1), while unconscientious people swear a lot and speak loudly (Rules 19 and 20). Our models contain many additional personality cues which aren't identified through a typical correlational analysis.</Paragraph>
  </Section>
class="xml-element"></Paper>