Predicting personality from text is notoriously difficult. While we might feel we can "read" someone through their writing, the actual signal in language is often weak and inconsistent. Most current AI models try to predict broad "Big Five" traits (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism) directly from text embeddings. But is a single label enough to capture the complexity of a human life narrative?
A new paper, FACET, argues that we should be looking deeper. By decomposing broad personality domains into their constituent "facets" and using contrastive learning, the authors have developed a framework that not only improves prediction accuracy but also provides a more granular look at how personality manifests in speech.
TL;DR
FACET is a three-phase framework that estimates Big Five personality traits from life narratives by decomposing broad domains into specific "facets" using supervised contrastive learning. It achieves state-of-the-art results (notably an $R^2$ of 0.328 for Openness) by combining deep language models with traditional psycholinguistic features through an attention-based pooling mechanism.
What Problem Does This Paper Address?
Personality computing faces two major hurdles. First, the "signal-to-noise" ratio in text is low; personality is a subtle background hum, not a loud shout. Second, most models treat the Big Five domains as monolithic blocks. In reality, a domain like "Extraversion" is composed of distinct facets like Warmth, Gregariousness, and Assertiveness.
Existing methods typically just concatenate Large Language Model (LLM) embeddings with psycholinguistic features (like word counts for "emotional" or "social" words) and hope for the best. The authors of FACET ask: Can we recover more signal by explicitly modeling these sub-facets and being more deliberate about how we combine different "views" of the text?
Methods
The FACET framework operates in three distinct phases, as illustrated in the architecture diagram below:
- Modality Encoding: The model takes two "views" of an interview transcript. It uses a frozen Mistral-7B model to capture deep semantic context and a specialized transformer to encode psycholinguistic features (LIWC and NRC lexicons).
- Facet Contrastive Learning: This is the core innovation. Instead of predicting the domain score immediately, the model learns six separate embeddings—one for each facet of a Big Five domain. It uses a supervised contrastive loss, which forces the model to cluster participants with similar facet scores (e.g., two people high in "Adventurousness") together in the embedding space while pushing dissimilar ones apart.
- Facet Pooling & Prediction: Finally, an attention mechanism decides which facets are most relevant for the final domain-level prediction. This recognizes that some facets (like "Intellect") might be more visible in a life narrative than others (like "Dutifulness").
The researchers also compared joint training (optimizing everything at once) against staged training (learning facet embeddings first, then freezing them to train the final predictor), finding that the end-to-end joint approach was superior.
Datasets
The study utilized the St. Louis Personality and Aging Network (SPAN) dataset.
- Participants: 1,332 older adults (ages 55–64).
- Input: Life-narrative interviews where participants recount key life events (averaging ~2,500 words per transcript).
- Ground Truth: Scores from the NEO-PI-R, a gold-standard personality inventory that provides both domain and facet-level scores.
Key Insights and Ideas
- Decomposition is Power: Breaking domains into facets provides a "structured" representation that helps the model navigate the weak signal in text. It also allows the model to provide facet-level predictions, which are much more useful for psychological assessment than a single broad score.
- Attention over Averaging: Simple averaging of features assumes every part of a story is equally telling. The attention-based pooling proved that letting the model "choose" which facets to focus on for a specific person leads to lower error.
- The "Openness" Advantage: The model performed exceptionally well on Openness to Experience. This suggests that the way people describe their life stories—the complexity of their thoughts and the variety of their experiences—is a very strong signal for this specific trait.
- Joint Training Wins: End-to-end gradient flow is crucial. Allowing the final regression error to "shape" the facet embeddings during training outperformed the more traditional staged approach.
Key Results
- Top Performance: The model reached an $R^2$ of 0.328 and a Pearson correlation of 0.596 for Openness, significantly outperforming baselines like BERT-based MLPs or standard psycholinguistic regressions.
- Facet Accuracy: Facet-level classification reached an AUC of up to 0.740 (for the "Ideas" facet of Openness).
- Ablation Success: Attention pooling yielded the lowest error across every personality domain compared to mean or concatenation pooling.
- Modality Synergy: Combining LLM embeddings with psycholinguistic features specifically improved results for Conscientiousness and Agreeableness, proving that "old-school" word counts still have value alongside modern transformers.
Limitations and Future Work
While promising, the authors note several limitations:
- Demographic Narrowness: The SPAN dataset consists of older adults from a single region. Personality expression in text likely changes across different age groups and cultures.
- The Neuroticism Gap: Like many models before it, FACET struggled with Neuroticism ($R^2 \approx 0$). This trait may be more internal and less likely to surface in the specific context of a life-narrative interview.
- Median Splitting: To use contrastive learning, the authors turned continuous scores into binary "high/low" labels, which discards some nuanced information.
- Future Directions: The authors suggest that fine-tuning the LLM (rather than using a frozen one) and validating the model on more diverse datasets are the next logical steps for this research.
Paper Citation: Arumugam, B., Parekh, P., Oltmanns, J., & Gupta, M. (2024). FACET: Facet Aware Contrastive Embeddings for Traits — Big Five Estimation from Life Narratives. arXiv:2611.23560v1.
Summary updated