Personality, Gender, and Age in the Language of Social Media: The Open-Vocabulary Approach

H. Andrew Schwartz, Johannes C. Eichstaedt, Margaret L. Kern, Lukasz Dziurzynski, Stephanie M. Ramones, Megha Agrawal, Achal Shah, Michał Kosiński, David Stillwell, Martin E. P. Seligman, Lyle UngarPLoS 2013 · DOI

Abstract

We analyzed 700 million words, phrases, and topic instances collected from the Facebook messages of 75,000 volunteers, who also took standard personality tests, and found striking variations in language with personality, gender, and age. In our open-vocabulary technique, the data itself drives a comprehensive exploration of language that distinguishes people, finding connections that are not captured with traditional closed-vocabulary word-category analyses. Our analyses shed new light on psychosocial processes yielding results that are face valid, tie in with other research, suggest new hypotheses, and give detailed insights. To date, this represents the largest study, by an order of magnitude, of language and personality.

Introduction

For decades, quantitative "lexical" psychology relied on closed-vocabulary tools like Linguistic Inquiry and Word Count (LIWC). These tools count how often people use words from a priori categories (e.g., "positive emotion" or "social process"). While effective, this approach is limited by the researchers' own preconceptions—you can only find what you are looking for. Furthermore, previous studies were often hampered by small sample sizes, typically involving only hundreds or a few thousand participants.

This research addresses the need for a data-driven, open-vocabulary method that can scale to the "Big Data" era of social media. By leveraging a massive dataset of 700 million words from 75,000 Facebook volunteers, the study pioneers an approach where the data defines the categories, uncovering unexpected psychological markers that traditional lexicons might miss.

Methods: Differential Language Analysis (DLA)

The authors introduced Differential Language Analysis (DLA), a three-step pipeline designed to extract and analyze linguistic features at scale.

The three-step Differential Language Analysis (DLA) pipeline: feature extraction, correlation analysis, and visualization.
  1. Feature Extraction: Instead of using a pre-defined dictionary, the researchers extracted all words and phrases (1- to 3-grams) using an emoticon-aware tokenizer. They also employed Latent Dirichlet Allocation (LDA) to automatically cluster words into 2,000 "topics" (e.g., a "school" topic containing homework, class, teacher).
  2. Correlational Analysis: Ordinary Least Squares (OLS) regression was used to identify language features that distinguish groups. The researchers controlled for age and gender to isolate the unique effect of personality traits, applying a stringent Bonferroni correction to ensure statistical significance.
  3. Visualization: A novel word cloud technique was developed where word size represents the strength of correlation with a trait, rather than just raw frequency.

Datasets

The study utilized the MyPersonality Facebook Dataset, comprising 15.4 million status updates from 74,941 volunteers. Each volunteer provided psychometric data by taking a standard 20- to 100-item Big Five personality test (NEO-PI-R). This scale—encompassing 700 million instances of words, phrases, and topics—is an order of magnitude larger than any previous study of its kind.

Key Insights and Psychological Mapping

The open-vocabulary approach captured nuances that traditional tools missed. For example, while traditional tools might show that neurotic individuals use more "negative emotion" words, DLA revealed they specifically use the phrase "sick of" and the word "depressed."

Word cloud showing language features significantly correlated with male users, including topics like sports, gaming, and the "possessive my" (e.g., "my_wife").
  • The "Possessive My": One of the most granular findings was that men use the possessive "my" significantly more often when mentioning their "wife" or "girlfriend" than women do when mentioning their "husband" or "boyfriend."
  • Stability through Activity: Emotionally stable individuals (low neuroticism) frequently mentioned social sports and life activities (basketball, snowboarding, church, meetings). This suggests that an active, socially integrated life is a strong marker of emotional stability.
  • Introverts and Interests: The study found a robust correlation between introversion and interests in Japanese media (anime, manga, pokemon) and Japanese-style emoticons (e.g., ^_^).

Key Results

The open-vocabulary models achieved state-of-the-art predictive accuracy:

  • Gender Prediction: The model achieved 91.9% accuracy in predicting gender based solely on language, outperforming previous benchmarks of 88.0%.
  • Personality Correlation: Models reached a correlation ($R$) of 0.42 for some traits. In psychology, $0.30$ to $0.40$ is often considered the "correlational upper limit" for predicting behavior, suggesting these models are hitting the theoretical ceiling of linguistic personality prediction.
  • Information Density: Models using automatically generated topics and phrases significantly outperformed those using traditional, manually-curated LIWC categories across all five personality traits.

Limitations and Future Directions

While groundbreaking, the study notes several limitations. The dataset consists of Facebook users from 2009–2011, which skews younger (mean age 23). Additionally, language on Facebook is performative and social; these markers might not generalize perfectly to other contexts like professional emails. Finally, the study identifies correlations, not causes. Future work could apply DLA to other fields, such as predicting health outcomes or analyzing ideological shifts in political science.

Summary updated

Abstract

We analyzed 700 million words, phrases, and topic instances collected from the Facebook messages of 75,000 volunteers, who also took standard personality tests, and found striking variations in language with personality, gender, and age. In our open-vocabulary technique, the data itself drives a comprehensive exploration of language that distinguishes people, finding connections that are not captured with traditional closed-vocabulary word-category analyses. Our analyses shed new light on psychosocial processes yielding results that are face valid (e.g., subjects living in high elevations talk about the mountains), tie in with other research (e.g., neurotic people disproportionately use the phrase 'sick of' and the word 'depressed'), suggest new hypotheses (e.g., an active life implies emotional stability), and give detailed insights (males use the possessive 'my' when mentioning their 'wife' or 'girlfriend' more often than females use 'my' with 'husband' or 'boyfriend'). To date, this represents the largest study, by an order of magnitude, of language and personality.
BibTeX
@inproceedings{schwartz2013personality,
  author = {H. Andrew Schwartz and Johannes C. Eichstaedt and Margaret L. Kern and Lukasz Dziurzynski and Stephanie M. Ramones and Megha Agrawal and Achal Shah and Michał Kosiński and David Stillwell and Martin E. P. Seligman and Lyle Ungar},
  title = {{Personality, Gender, and Age in the Language of Social Media: The Open-Vocabulary Approach}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pone.0073791},
}
Figure