Studies that correlate personality with language features use homogeneous data sets from a single domain, such as Facebook posts, emails from one company, and student essays from one university. Such single-domain correlations may not generalize well to other domains. Therefore it is important to study which language features are associated with personality across multiple domains. This paper reports preliminary results from the rst such cross-domain study of correlations between personality and language usage.
Personality traits are consistent patterns in a person's behavior over time|particularly behavior that observers consider when forming an opinion about how an individual's behavior di ers signi cantly from others. A prevailing model of human personality, the Five Factor Model, places such behavior in ve dimensions: extraversion, agreeableness, conscientiousness, neuroticism, and openness.
Language usage tends to reveal a lot about someone's personality. The advent of computer technology, particularly digital storage and retrieval of text allows us to examine language usage. When relevant, such as in e-mail exchanges, speech acts may predict personality (e.g. the disagreeable person is apt to repeat demands without o ering a variety of other speech acts), and punctuation and word sentiment certainly do.
Word frequency (bag-of-words counts) along with overall stem and word counts comprise some of the most intuitive and common features extracted from text. Since word usage is quite context dependent, we are interested in examining aspects of language usage that are less so. Part of speech n-grams preserve information about how a speaker is using language while decoupling from speci c words, which are very context dependent. Also n-grams combining both words and part of speech present a compromise between pure word usage and grammatical usage.
We worked with two sets of participants. The rst set is
a new corpus that we collected ourselves: group of 49 web
forum users to whom we administered an personality test
consisting of 50 items from the IPIP [
We extracted various POS n-grams, and sometimes
hybrid POS and word n-grams. Table 2 shows a few examples
of these features and the text that underlies them. The
hybrid features provide context about word usage that simple
word counts lack. To extract the features, we rst tokenized
each participant essay and then extracted the features of
interest; when possible used pre-existing tools. Although
the statistics computed were straightforward, we chose to
use standard, well-tested statistics libraries to avoid errors.
The POS tagger we used was an implementation of that
presented in [
The most populous sample (the Essays corpus) has 2588 participants. We chose a number 1/5th our sample size: 517. We took the 517 most frequent features F in the Essays corpus (the corpus with the most participants) and ignored the rest. Then we extracted F from the Forum texts (49 participants) as well. Of the 517 features, we show here only the ones with p < 0:1 for both corpora. 4.
To determine the independent relatedness of these features to the personality dimensions, we computed the Pearson correlations, Table 1(c, e), between feature frequencies, normalized to document length, and scores in the given personality dimension, Table 1(a). The p-values tell us the probably of the null hypothesis. An encouraging aspect of this early result is that far more features of interest are related to Conscientiousness than any other personality dimension. This suggests a close relationship between usage of these language features and the speaker's personality in the Conscientiousness dimension. It is di cult to imagine an unrelated process that would cause such a signi cant difference. (c) (d) Essays corr p-val (e) (f)
Forum corr p-val
For the Essays corpus, the e ect sizes are small, whereas for the Forum corpus they are consistently larger, Table 1(c, e). The sparsity and variance of the Essays features may be constraining their predictive impact; Essays features have an average frequency of 4 whereas Forum features have an average frequency of 248. The Forum texts are generally much longer than the Essays, so that constraint is removed. 5.
It may be possible to extend this work by including coarsegrained parts of speech (e.g. noun phrases) extracted by chunking tools. Further examination of additional corpora may establish the generalizability of our language features to a variety of populations. Also compelling explanations of why particular POS n-grams are indicative of personality would be of great interest in directing the exploration of new text features useful for personality prediction. Finally, the sparseness of some corpora encourages analysis of features measured by assigning f1; 0g when fpresent; absentg, a practice that is sometimes useful when working with sparse features. 6. 7.
We extend grateful thanks to James W. Pennebaker for making this research possible by sharing the essays and personality scores. 2.214 2.009 -2.994 -2.415 1.307 1.624 forum forum essay essay forum essays forum forum essay open open open cons cons extra extra agree neur see if there's the possibility to get more calories I have the tendency to tip a minimum not in the mood to do it also having to move out I am trying to sleep because I came down with I don't snack often if i'm trying to lose weight should have started writing at a time that was easier high end CPU is a little bit of a waste for gaming