Você está na página 1de 16

Taking Care

Of the Linguistic Features of Extraversion


Alastair Gill and Jon Oberlander University of Edinburgh

Extravert or Introvert?

Jim Carrey: Introvert?

Albert Einstein: Extravert?

Which are they ... and where are you?

12/2/2013

Gill & Oberlander

Personality meets HCI

Reeves and Nass had subjects solve a problem and then discuss solution with the computer (via a textual interface) Language variables manipulated to provide different system personalities:

Dominant:
Always goes first Strong language Uses assertions and commands Indicates high confidence
12/2/2013

Submissive:
Goes second Unassertive language Uses questions and suggestions Indicates low confidence

Gill & Oberlander

Eysenck's Three Factor PEN model

Extraversion

Neuroticism

social interest and positive affect "arousal level in the cortex" (lower levels = more extravert) Response to stressors and negative affect "activation thresholds in the limbic system" (lower thresholds = more neurotic) Aggressiveness, individuality "testosterone levels?" (higher levels = more psychotic)

Psychoticism

What are the implications for language?

12/2/2013

Gill & Oberlander

Personality and Language ... so far

Work has focussed mainly on Extraversion in speech What has this found?
Talk more in discussions (Carment et al., 1965) Distinguished by speech cues (Scherer, 1979) Specific (eg. syntactic category) and less specific (style) features related to lexical choice proposed (Furnham, 1990) LIWC content analysis program used with transcriptions to look at warmth and dominance facets of Extraversion (Berry et al., 1997)

12/2/2013

Gill & Oberlander

Textual Personality

Relatively little work on written linguistic personality


Written diary entries analysed using LIWC (Pennebaker and King, 1999)

Can we replicate Pennebaker and King's findings? How else is personality embodied in text
Which linguistic features are most important for personality? What can more sophisticated computational linguistics techniques reveal?

12/2/2013

Gill & Oberlander

Our personality corpus

105 subjects generating two texts each Each completed the Eysenck Personality Questionnaire Each then composed two emails:
"To a good friend whom they hadn't seen for quite some time" One concerned past activities over the previous week The other concerned planned activities over the next week.

Each message took around 10 minutes to compose and submit by HTML form. The resulting 210 texts contain 65,000 words.

12/2/2013

Gill & Oberlander

Experiment 1: replication and text analysis Factor analysis


Selection of LIWC items for principal components analysis used the same criteria as Pennebaker and King Four factors derived which are essentially similar to theirs, with minor variations in factor loadings

Multiple regression analysis


Variables which showing a small correlation with personality type and topic independence were entered for:
LIWC MRC Psycholinguistic Database (used to compute mean scores for verbal frequency, written frequency, concreteness, age of acquisition, etc.)

12/2/2013

Gill & Oberlander

Interpretation

Extraverts do write more


This echoes earlier findings on speech

And they write more loosely (with fewer exact number expressions)
For eg. several vs five Supports Furnhams hypothesis

And we also find they write less concretely


Eg. Furniture vs chair

But:
The variance explained is not as great as for P and N Correlations not as strong as found by Pennebaker and King

12/2/2013

Gill & Oberlander

10

Experiment 2: NLP techniques

Use simple bigram techniques to


take more of a word's context into account

Select high-E and low-E subcorpora by taking texts from subjects with E-score > 1 s.d from mean (cf. Dewaele and Pavlenko, 2002):
21 High E versus 17 Low E 12,000 words versus 8,000

Generate bigram profiles ranked by Dunning's log-likelihood statistic (top 50 bigrams with frequency >= 2, p < 0.001). Calculate relative frequency ratios (Damerau 1993) for bigrams common to both subcorpora

12/2/2013

Gill & Oberlander

11

On the surface

The gross features are perhaps the most intuitive in their representation of the Extraverts or Introverts. For example, [<START> hi], the <START> marker followed by hi, was unique to Extravert texts;
message-initial hi.

By contrast the more formal [<START> hello] was found solely in Introvert texts Use of punctuation also differs between the two groups:
Extraverts preferring multiple exclamation marks [! !], and solely using multiple full stops [. .] as in the elliptical (...) features of informal style, and looser use of language.

12/2/2013

Gill & Oberlander

12

Quantification

Introverts tend to show preference for a greater use of quantifiers


[a lot], [a few] and uniquely [all the], [one of], [lots of] and [loads of]

Extraverts show a preference for [a bit] and uniquely use [couple of]. Not only does this demonstrate an Extravert tendency to be looser and less specific, it also apparently reveals a tendency towards exaggeration on the part of the Introvert.

12/2/2013

Gill & Oberlander

13

Valence

Bigrams containing negations were used significantly only by Introverts, as in [i dont] and [dont know]
(indeed [i dont] is the bigram with most frequent use of i)

Similarly, the Extravert preferences suggest a more positive, relaxed disposition:


[looking forward] and [forward to] (presumably as in looking forward to) [a good] [catch up] [take care]

12/2/2013

Gill & Oberlander

14

Ability and Modality

Personal views on capability are suggested by the different collocations with infinitival to.

Similarly, collocations with the verb be show a distinction in use of modal auxiliaries which has an effect on the projection of certainty.
For example, Introverts uniquely use the weaker [should be] Extraverts prefer the stronger predictive [will be], and contracted form [ill be] (i will be).

Extraverts demonstrate ability with want, need, and able (to) Introverts more timidly and tentatively are [trying to] or [going to]

12/2/2013

Gill & Oberlander

15

What about the P and N dimensions? Bigrams in 3D

High-E use ... and !!


Both associated with High-N; But one goes with High-P; the other with Low-P

High-E use [take care]


This is Low-N, Low-P

Low-E use [, but] and [, because]


One is High-N, High-P; The other is Mid-N, Mid-P.

12/2/2013

Gill & Oberlander

16

Summary

CASA tells us that linguistic personality might matter Simple techniques can confirm known linguistic features of extraversion and uncover new ones Applications:
Interface agents Personality Language Checker

Future work:
Test sensitivity of readers to personality features Investigate feature generalisability

12/2/2013

Gill & Oberlander

Take care!
17

Você também pode gostar