Embedded Personalities: Word Embeddings and the “Big Five” Personality Model

Oliver Müller, Stefania Degaetano-OrtliebNAACL 2025

TL;DR

This paper introduces a novel, unsupervised method for analyzing personality traits by using word embeddings to represent adjectives associated with the Big Five personality model. By mapping these adjectives into a multi-dimensional space and applying dimensionality reduction and clustering, the researchers successfully recreated the Big Five dimensions, offering a scalable alternative to traditional self-assessment methods, especially for historical or literary texts.

What Problem Does This Paper Address?

The Big Five personality model (OCEAN: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) is a widely accepted framework for understanding personality. Traditionally, these traits are measured using self-assessment questionnaires, which rely on participants' subjective reports. While effective, this approach has limitations: it can be prone to introspection illusions (self-perception biases), assumes linear relationships between traits, and is not applicable to contexts where self-reporting is impossible, such as analyzing characters in literature or historical documents. The authors aim to address this by exploring an unsupervised method that leverages word embeddings to capture the semantic relationships between personality-descriptive adjectives, thereby bypassing the need for self-reported data and enabling analysis in new domains.

Methods

The core of this research lies in applying word embeddings to personality analysis. The authors used a pre-trained Word2Vec model (Google-News-300) to generate 300-dimensional vector representations for 100 adjectives associated with the Big Five personality dimensions. These adjectives were carefully selected, with 20 allocated to each of the five dimensions, drawing from existing lists and supplementing with adjectives generated by ChatGPT-4o to ensure broad coverage.

To make these high-dimensional vectors interpretable and to identify underlying structures, two key techniques were employed:

  • Principal Component Analysis (PCA): PCA was used to reduce the dimensionality of the word embeddings to a more manageable three dimensions. This process captures the most significant variance in the data, allowing for visualization and analysis of the relationships between adjectives. The choice of PCA was motivated by its ability to maintain global structure and provide interpretable linear projections.
  • K-means Clustering: Following PCA, K-means clustering was applied to group the adjectives into distinct clusters. This unsupervised method aims to identify natural groupings within the data without prior labels, allowing the researchers to see if the word embeddings naturally form the Big Five dimensions.

A silhouette analysis was conducted to determine the optimal number of clusters, which confirmed that five clusters provided the best separation, aligning with the Big Five model.

Datasets

The study utilized a pre-trained Word2Vec model (Google-News-300) for generating word embeddings. This model was chosen for its extensive training on a large and diverse corpus, ensuring broad coverage of personality-descriptive terms.

The adjectives used for analysis were compiled from:

  • Existing lists, including those from Tupes and Christal (1961).
  • 50 additional Big Five adjectives generated using ChatGPT-4o, specifically briefed to ensure even distribution across the 10 aspects of the Big Five dimensions.

The study did not introduce a new benchmark dataset but rather applied its methodology to existing linguistic data represented by word embeddings.

Key Insights and Ideas

  • Unsupervised Validation of Personality Dimensions: The study demonstrates that word embeddings can successfully capture and recreate the semantic structure of the Big Five personality dimensions without relying on traditional self-assessment data. This provides an unsupervised validation of established personality constructs.
  • Bridging Computational Linguistics and Psychology: The research offers a novel way to bridge computational linguistics (word embeddings) with psychology (personality models), opening up new avenues for interdisciplinary research.
  • Scalable Analysis for New Domains: The method's unsupervised nature makes it highly scalable and applicable to domains where traditional survey methods are not feasible, such as analyzing literary characters, historical documents, or social media text.
  • Visualizing Semantic Relationships: The use of PCA for dimensionality reduction allows for intuitive visualization of how personality-related adjectives cluster together, revealing semantic relationships and potential overlaps between traits.
  • Identifying Overlaps and Distinctions: The visualization clearly shows separations between dimensions (e.g., Neuroticism vs. others) and also highlights overlaps (e.g., Extraversion and Agreeableness), providing nuanced insights into the structure of personality.

Key Results

  • The study successfully recreated the Big Five personality dimensions using word embeddings, PCA, and K-means clustering.
  • A silhouette analysis indicated that five clusters provided the optimal separation, supporting the hypothesis that the Big Five dimensions are reflected in the adjective embeddings.
  • Statistical analysis (one-way ANOVA) confirmed statistically significant differences between the clusters across all three principal components (PC1, PC2, PC3), indicating that the trait-associated adjectives form distinct groups in the PCA space.
  • The visualization of adjective clusters revealed interpretable patterns, with dimensions like Neuroticism showing clear separation, while others like Extraversion and Agreeableness exhibited some overlap.
  • The contribution of personality aspects to PCA showed that PC1 captures the opposition between positive and negative emotional traits, PC2 distinguishes between withdrawal and politeness, and PC3 differentiates structured/orderly traits from spontaneity.

Limitations and Future Work

Limitations:

  • Context-Dependent Meanings: Static word embeddings may struggle to account for adjectives with context-dependent meanings, potentially leading to misinterpretations.
  • Computational Demands of Advanced Embeddings: While contextualized embeddings (like BERT) could address context sensitivity, they come with significantly higher computational demands.

Future Work:

  • Validation with Randomly Sampled Adjectives: To ensure that the observed clustering is not an artifact of the preselected adjective lists, future work will involve comparing results with randomly sampled adjectives.
  • Expanding Adjective Lists: Incorporating a broader and more comprehensive set of adjectives could enhance the granularity of trait analysis and potentially reveal more nuanced personality profiles, especially in literary texts.
  • Domain-Specific Embeddings: Calculating embeddings directly from domain-specific textual corpora (rather than relying solely on general pre-trained models) could lead to more accurate and context-aware personality analysis.
  • Integrating Contextualized Embeddings: Future research could explore integrating contextualized embeddings to better handle sentence-level nuances and improve the analysis of complex and nuanced textual data.
  • Exploring Other Visualization Techniques: While PCA was effective, exploring other dimensionality reduction techniques like t-SNE or UMAP could offer alternative perspectives.
  • Temporal Analysis: Visualizing how personality traits evolve over time within narratives or rhetorical contexts by segmenting texts temporally.
  • Broader Applications: Applying the methodology to analyze semantic shifts in adjectives over time, trace diachronic semantic changes, and investigate rhetorical styles in historical correspondence.

Summary updated

BibTeX
@inproceedings{muller2025embedded,
  author = {Oliver Müller and Stefania Degaetano-Ortlieb},
  title = {{Embedded Personalities: Word Embeddings and the “Big Five” Personality Model}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure